Changelog¶
Unreleased¶
Fixed: the search for early paths takes a reflection point within its tolerance for planes of the next reflection’s plane (a ten-millionth of the model’s largest extent, 1 mm in a model 10 km wide) as the corner of the two, and such a path met each plane at the angle of its legs to that corner, which can lie as far from where the path meets it; a reflection with both legs shorter than the tolerance met its plane at grazing incidence, and with edge diffraction the edge fade dropped a joined path with all its terms. A joined path now takes the points the image method places, where that path can be, and every reflection meets its plane at the angle the source’s image gives. Against the same scenes in a model 200 m wide, a source and a listener 0.1 mm before a wall and 0.5 mm apart had been 20.7 dB too low, a source 0.3 mm from both arms of a bent barrier and the ground up to 2.6 dB, and with edge diffraction one 0.1 mm from both arms 1.2 dB at 63 Hz.
Added:
simulation.wave.walls: fitted(Fitted walls in the editor) puts the wave solver’s walls where the model has them, its faces cut exactly through the grid’s cells, instead of on the nearest cell faces: in a box along the grid the modes come within 0.05 % of the exact ones where walls on cell faces are up to 5.2 % off, and the level in front of a rigid wall within 0.04 dB where they are up to 1.8 dB off. It takes the standard stencil, indoors and in still air, refuses models with cracks or faces laid on one another (naming the faces), and has measured limits at steep walls, concave edges, free edges and absorbing walls (Low frequencies); saved scenes now carrywalls.Fixed: in a very large model, a corner of two reflections (points within the early search’s tolerance for planes of each other, a ten-millionth of the model’s largest extent, 1 mm in a model 10 km wide) was judged by that tolerance: a path into it went on through one of its planes unless the points before and after it lay farther than that from the plane, and it took a side of a plane only from such a point. A source 0.3 mm from both arms of a bent barrier in a model 10 km wide had a path through the barrier, with the early level behind it up to 42 dB above that of the same scene in a model 200 m wide, and one 0.3 mm from two walls of a building or from both barriers of a T reflected off the roof from under the eave or off the T’s stem from its bare side (up to 4.3 dB in a band). Both now take the least distance at which a ray meets a face.
Fixed: with Doppler, paths off the same surfaces over the same edge were paired from place to place in the order of their delays, so where one more came among them (as round a barrier’s end with edge diffraction) a path could go on as another whose delay was tens of milliseconds shorter, and its sound was heard going back in time for a moment: a click 40 dB above the sound around it, with a slow thump either side, 0.76 s into the walk of the
railwayexample’s train heard attsi_7m5(once for each of its bogies). Each path now goes on as the one whose delay carries on from its own; in that train’s walks to the example’s four listeners, 89 of the 502 paths followed jumped so, and none does now.Fixed: with edge diffraction, whether the end of an edge stands on a plane (a barrier’s end on the ground) is read from the plane’s faces at the end, no longer a ten-thousandth of the model’s size past it (1 m in a model 10 km long). With the ground ending 0.5 m past a barrier’s end in such a model, a listener round the end had 10 early paths in place of 8 and up to 5.0 dB less (at 125 Hz, with
first_order).Fixed: a free-standing panel or a partition between two rooms in line with a wall shared the wall’s plane in the model or had one of its own by the way the wall faces, so with edge diffraction a room and its mirror image gave different early levels (up to 1.9 dB at 63 Hz in the cases measured), and an inside corner of such a wall could count as a diffracting edge. The panel now lies in the wall’s plane whichever way the wall faces (with walls that face both ways in that plane, in the plane of one of them, as before), and such a corner is no edge.
Changed:
simulateand the editor work the draws of the materials’ spread out in worker processes, with the same results (on an Intel Mac the admittances’ fits may differ in their last digits, as they may there from one fit to the next): as many as the run has threads (--threads,simulation.performance.threads) and cores, one for each 100 draws of all the responses at most, and no more than half the computer’s memory holds. On four cores the 400 draws of theconcert_hallexample took 21 s where they took 72 s, and those of thestreetexample 163 s where they took 609 s. In Python,material_spreadandwith_spreadtakeworkers(1 by default, 0 for every core).Fixed: the faces of a plane at 45 degrees to two axes with air on both sides (a free-standing panel, a partition, any surface of a model that is not closed) could face opposite ways, by the rounding of their corners, and be two planes of the model. With
first_order, a room with three pieces of such a panel had 1,097 early paths in place of 1,205 when it was turned by 180 degrees, and its early level changed by up to 1.1 dB (at 63 Hz).Fixed: with edge diffraction, where a path reflects off two surfaces next to the line they meet at, the leg between them can be shorter than a millionth of the model’s size, and the terms that make reflections fade at edges then lost their paths over edges (the ray that looks whether such a leg is clear started past its end). With a source 2 mm from both arms of a barrier bent by 30 degrees in a model 10 km wide, a listener behind an arm had 1 of its 6 paths and up to 7.1 dB less (at 125 Hz, with
first_order).Fixed: with edge diffraction, a source so near two planes at once that its images in the two lie within a millionth of the model’s largest extent of each other (nearer both than 0.35 millionths where they meet at a right angle, 3.5 mm in a model 10 km wide: by a building’s eave or upright corner, or by the ground and the plane of a wall past the wall’s end) had the reflection off one plane taken for the reflection off the other where only one of the two reaches the listener: the early level was 0.9 to 2.6 dB off in the cases measured. Levels also move, by up to 0.44 dB in the cases measured, where two sequences of reflections off parallel planes have one image (off a wall and then a screen exactly halfway to the opposite wall, and off the screen and then that wall): the reflection off one sequence’s planes is no longer taken for the other’s.
Fixed: with
late.deviceset togpu,bothorauto, the late part came out low where the source or a listener lay right by a surface. In the cases measured: up to 3.2 dB for a source within a millionth of the model’s size of one, 6 to 7.5 dB for a source two millionths from both sides of an inner edge of 30 degrees (0.4 to 0.5 dB in a right-angled corner), and 2 to 6 dB for a listener 3 to 12 millionths over a floor that scatters. The CPU now traces such a run and the run says so (which runs).Fixed: a measured directivity (a CSV or SOFA file) whose directions lie equally near a point of the 5 degree grid it is interpolated on, as halfway between the directions of a table given every 10 degrees, took four of them by the last digits of a sum, and a direction given more than once (-180 and 180 degrees of azimuth, the rows at a pole) counted once for each row: a table that is the same left and right gave a pattern whose sides were up to 1.1 dB apart. Directions equally near a point are now taken alike and a direction given more than once counts once (Scenes), and directions that are not finite or are (0, 0, 0) are refused. The examples’
wedge.csvandbox.csvmove by up to 1.1 and 1.4 dB at single directions (the box’s 8 kHz band is about 0.3 dB down in most directions), and with them the levels ofconcert_stageby up to 0.5 dB at 8 kHz and 0.2 dB(A), those ofstadium_concertby up to 0.8 dB at 8 kHz and 0.2 dB(A), and the lakeside system’s 8 kHz band at the audience by 0.1 dB; a table given on the grid itself moves in its last digits only.Fixed: near the line where two walls meet in a very large model, a reflection off one wall went on past its edge over the other (by the search’s tolerance for planes, a ten-millionth of the model’s largest extent, times a factor of the face’s shape, or where its path met the other wall within that tolerance of the first), a reflection point within the tolerance of another plane reflected off that one too though the path did not go on to it, or ended the path beside its faces, and two reflections whose images fell in one step of the tolerance counted as one: 16 mm from a bend in a model 10 km wide, the early level was off by up to 6.7 dB in a band at 32 of 96 places 7 m away for a bend of 170 degrees, and by up to 4.2 dB at 4 for one of 91. A face now ends at its edges to within a billionth of the model’s diagonal.
Fixed: of two different early paths with one image of the source that reflect off different surfaces, such as those between two rooms as deep either side of a partition (off the partition and the source’s room’s end wall, or off the other room’s end wall and the partition), one was left out: there a listener had 63 of 70 paths up to the third order, and up to 2.2 dB less in a band.
Fixed: a path exactly through an outer edge (a roof’s edge over an end wall, a building’s upright corner, the edge where two barriers meet) stood as a reflection off both faces for a source and a listener the edge sees in the same direction across it, with
early.method: image_sourcesand, where a ray meets other faces in the two planes one after the other (a taller building’s wall in the plane of the end wall), with the default search: 1.5 m above the level of a rigid hall’s roof, midway between its end walls, the level was 5.3 dB higher at 63 Hz (first_order, image sources) than 0.1 mm beside. The validation’s long models had such paths off their far ends, so a few of its numbers for image sources move (docs/validation.md).Changed: a run’s results no longer depend on how many threads the linear-algebra library behind numpy is given (as many as the computer has cores, unless
OPENBLAS_NUM_THREADSsays otherwise): the package adds its long sums itself, each in a fixed order (acoustic_spaces.products; what stays with the library is in Simulation). Results move once in their last digits (in the examples: parameters by at most 1e-9 of their value, sound files by 3e-10 of the file’s peak, one response’s standard errors by up to 5e-5), and the examples’ runs take up to a quarter less time at 30 to 50 % of the CPU time; a rotor’sbpmbroadband noise takes 23 % more.Fixed: a path that reflects more than once at one point (a room’s corner sends sound back along the line it came on) took a surface with no leg before or after it at a grazing angle: in the bands that add in pressure its pressure came out turned over (with another factor and energy where that surface is a ground with an impedance). For a source and a listener on one line from the corner the early level was 0.7 dB off at 63 Hz in a room 7.3 by 5.1 by 3 m, and 1.6 dB (2.3 dB on grass) at the foot of two barriers that meet at a right angle, with the default settings.
Added: a solver for sound that passes close over a whole row of roofs and barriers in still air (
acoustic_spaces.engine.rows: the parabolic equation marched over the row, with the terms of its rigid edges and, for a barrier alone on a rigid ground or roof, the field of the strip it makes with its image,acoustic_spaces.engine.strip), within 0.38 dB of a wave calculation near grazing incidence in free space and 0.65 dB there over roofs on a rigid ground, 0.74 dB up from a source at street level and 1.24 dB down to a listener there; between rigid screens on a rigid ground up to 3.4 dB low, a parapet lower than a quarter of a wavelength on a roof’s edge up to 1.3 dB off, and with an error for the ends it cannot hold, such as one almost straight under the top of a wall (validation/rows.py). A way that crosses the row at a slant takes each edge at its slant (slants, up to 75°; one screen within 0.10 dB of the exact solution), and where an end’s way to the row is steep or an edge in a row is turned more than 60°, a row on a rigid ground whose edges all lie in one direction is solved across them and moved to the slant: two roofs on a rigid ground within 1.3 dB of the wave calculation at 30, 60 and 75°, three rigid screens there up to 8.2 dB low.Changed: outdoors, the sound over long barriers and rows of buildings in the vertical plane through a source and a listener comes from the solver for whole rows in the octave bands up to 500 Hz that add their paths in pressure (
simulation.early.rows: solver, the default;--rows, Barriers and buildings in the editor), in still air and, scaled by the weather’s effect, in wind and temperature gradients: against a wave calculation two flat roofs on a rigid ground are within 0.74 dB, where the paths over the edges are up to 10.7 dB off.rows: edgesgives the levels as before; coverage maps and walks keep the paths over the edges (Outdoor scenes).Fixed: a source or a listener counted as on a surface only within 1 µm of it whatever the model’s size, though a ray traced on the CPU meets no face nearer its start than 1e-9 of the model’s diagonal: twice that distance now counts where it is more, in a model more than 500 m from corner to corner. A listener 6 µm over the hard ground of a model 10 km wide, almost under its source, was 6 dB low without a message.
Fixed: the wave solver missed a wall through a plane of its cell centres or just past one, so that sound went through it, a closed room could be read as open and the 19-point run could grow without bound; a source or a listener exactly on a surface is now read a hair off it. Results change only where a surface comes within about 1e-9 of the model’s size of a cell centre, the source or a listener, or a link or line of sight meets a surface within 1e-10 cells of the line of one of its edges; surfaces about 1e-10 cells off the centres can still be placed wrong, up to a closed object read as air (Known issues).
Changed: an outdoor ground (a material with an
impedance) scatters nothing unless its material gives ascattering, where it took 0.1 like any material: the early level over grass is within 0.05 dB of the exact solution 20 and 100 m away (it was 0.3 and 0.5 dB off at 500 Hz), the ground dips in still air are deeper, and flat open ground alone gives no late part.scattering: 0.1on the ground gives the earlier results.Fixed: with
second_orderdiffraction, sound that just clears two edges in a row (a source and a listener in sight of each other over a flat roof, a thick wall or two barriers, or along a wall in line with both) was far too low, 15 to 26 dB across the bands a centimetre above a 10 m roof and 6.7 dB at 63 Hz half a metre above it: the path over both edges was left out there.first_orderhas no such path and stays low there.Fixed: with a source and a listener both exactly level with the tops of two or three barriers in a row, each top’s path stood for the whole direct sound, about 10 dB too high over two. The level is now that just above the tops, which
first_orderanddirecthave far too low there (they have no path over two edges).Fixed: just above a flat roof and a barrier whose top is level with it, the level changed by up to 8 dB with the source and the listener swapped and, past the barrier’s end, by up to 16 dB between a millimetre and a centimetre above their plane: a path over the barrier’s top was taken to reflect off the roof, whose plane it meets at the top itself, and the barrier’s term was taken on the wrong side of the reflection off the roof one way round.
Changed: with edge diffraction, the first reflections the rays look for include the faces of every diffracting edge the source stands in front of, which a ray misses where the source sees a face nearly edge-on or an obstacle hides it. Levels near such a face can move with every diffraction mode (by 0.2 dB at 63 Hz at one of eight positions in two cities with
first_order), and a whole city simulates up to about 15 % slower.Added: a coverage map gives the uncertainty of its values from the materials’ absorption uncertainty, an insertion loss and the long-term levels too:
coverage.spread: true,map --spreador the tick box under Uncertainty on the editor’s coverage map page trace the map twice more for each material that has one (the law of propagation of uncertainty, GUM 5.1.2).map --showand the map panel’s Show draw the standard uncertainty, an end of the 95 % interval or a one-sided bound.Fixed: a calibrated material given in third octaves keeps the octave values that were fitted where one of its thirds reaches 1 or what the surface lets through: the octave’s other thirds take up the rest.
Fixed: a run re-weighted for much more absorption than it was traced with (a draw of a wide absorption uncertainty, a large change in the editor) gave far too much late energy, up to absurd levels near an absorption of 1. The late part now follows Eyring’s decay, which stays near tracing again.
Fixed: the editor’s Run demo… asks to save a scene that has no file or has changes, so the window lists its sources to pick WAV files for instead of an empty table.
Added: the spread’s draws can go on until every result is stable to the first digit of its standard uncertainty, as the adaptive procedure of GUM Supplement 1 has them (JCGM 101:2008, 7.9):
simulation.uncertainty.draws: adaptive,--spread-draws adaptiveor Draws in the editor’s simulation settings, which also take a number (400 as before) or none.parameters.jsonsays whether the results were stable.Added: the spread gives each result’s one-sided bounds, for a limit on one side: the value it is at least and the one it is at most with a probability of 95 % (
at_least,at_mostinparameters.json, insimulate’s output and in the editor’s tips).simulation.uncertainty.one_sided,--one-sidedor One-sided bounds in the editor’s simulation settings give another probability, 90 % say.Added: a material’s page in the editor sets its absorption uncertainty (Absorption uncertainty: a standard uncertainty or a tolerance per octave band and its source, or ISO 12999-2’s reproducibility at a click), which only the scene file did. Typing a material’s coefficients there no longer drops an uncertainty the scene file gave it.
Fixed: typing one coefficient on a material’s page in the editor changes that cell only. A scattering value no longer takes the material off its database entry or file (which cost a built-in material the uncertainty a run takes for it), the other bands keep the decimals the table does not show, and a material given by its sound reduction alone takes a scattering value.
Fixed: a material that absorbs everything in some band (five built-in ones do) was left out of the spread in every band, without a word. Its other bands are now drawn, and the spread names the bands it holds (
heldinparameters.json, the output and the editor’s tips).Fixed: a run that cannot have a spread says what the spread needs (the editor too, which said nothing), at most 16 materials on the surfaces among it, and
simulateno longer announces a spread that no material on the surfaces gives.Fixed: the errors over edges that the spread lists as not counted follow the settings of the run (
simulate --diffraction ...), not those in the scene file.Fixed: the spread drew an absorption uncertainty given beside a ground’s impedance as if that ground reflected 1 - absorption, which its impedance decides. It is now left out, with a warning, and a draw no longer goes below what a material lets through.
Fixed: a response’s spread no longer depends on the other responses of the run, whose number moved its mean, uncertainty and interval within the draws’ noise; in a run of several responses, all but the first change by that much.
Fixed: a material copied with other absorption (
calibration.with_absorption) no longer counts as built-in, which gave it the built-in materials’ default uncertainty.Changed: the spread’s combined standard uncertainty is the root of the sum of the squares of its budget (
absorption,rays), where the standard deviation of draws that held the rays’ deviates was unsteady with 4 or 5 batches, and its mean is that of the absorption’s draws. Both move within the draws’ noise.Fixed: an end of the spread’s interval or a one-sided bound no longer passes what the parameter can be (a T30 from -0.18 s, a D50 up to 1.0003): it is held at that limit.
Fixed: on a material’s page in the editor, ticking Absorption uncertainty no longer rewrites a scattering entry given by name as numbers, the Source field keeps its caret after Enter, an opening takes no uncertainty, and a source reads “edited from …” once the kind or a value of the uncertainty it names is changed.
Changed: behind barriers and buildings in wind, the parabolic equation’s runs that differ only past their first obstacle share the march over the open ground before it, and the march steps eight rows of its field at a time: coverage maps of 90 to 96 points take 77 to 89 s on one core where they took 104 to 121 s, with the same results.
Changed: the wide-angle parabolic equation’s steps take up to four frequencies together, two to an SSE2 register: listeners the sound reaches steeply behind a barrier or a building take about 60 s on one core where they took 115 s, with the same results.
Changed: a walk’s places between those traced for their late part mix theirs when it is read instead of holding a copy for the whole run: a walk of 201 places through the
churchexample peaks at 336 MiB where it took 1,685 MiB, with the same results, and the memory estimate counts what is kept.Changed: on the CPU, the ray tracer finds the nearest of a run’s directions for an arrival in a table instead of going through them all and no longer clears its work arrays for every ray, and the wave solver works out each state’s term of a wall’s branches once a step instead of twice: the late part of the
concert_hallexample takes 9.8 s on four cores where it took 13.6 s (thelecture_hall’s 7.4 s where it took 9.5 s), and the wave solver’s run of theroom_modesexample up to the 500 Hz band 5.5 s where it took 6.8 s, with the same results.Changed: the draws of the materials’ spread share what does not depend on the draw (the paths’ counts on each surface, the early paths’ delay phases) and work out a draw’s octave echogram once for the parameters and the levels: with the faster fit below, 400 draws of the
concert_hallexample take 70 to 75 s on four cores where they took 157 to 158 s, with the same results.Changed: the fit of an admittance to a material’s absorption takes its forward differences a branch at a time: the simulation of the
shoeboxexample in third-octave bands takes 20 s on four cores where it took 30 s, with the same results (on an Intel Mac a fit changes with what else is in memory, as it did before: by 0.0000002 of absorption in a test).Changed: a run’s files are rendered on the threads its responses leave free, a format’s channels side by side as far as a quarter of the computer’s memory holds them: the files of the
shoeboxexample in third-octave bands take 3.6 s on four cores where they took 8.0 s, with the same results, and the run peaks at 1.1 GiB where it peaked at 0.5 GiB.Changed: the tests write every thread’s stack to the log when one has run for 20 minutes (pytest’s
faulthandler_timeout), to show where a hung job stands.Added: the spread covers each listener’s levels and speech results: the sound pressure level per octave band and A-weighted, the STI, the STIPA, U50 and, outdoors, the level relative to free field (
uncertainty.levelsinparameters.json, the A-weighted level’s and the STI’s insimulate’s output, the level’s in the editor’s tips).Fixed: Doc 29’s L_Amax ahead of or behind an airborne segment takes the installation correction at the extended path’s closest point of approach, as the SEL does, not at the segment end’s elevation: on Volume 3’s departures up to 1.0 dB higher beside the track past the cutbacks and up to 0.40 dB lower under it; the airport example’s listeners do not change.
Fixed: the Doc 29 table readers match a column by its whole number, so an NPD table whose distances run from 25000 ft down to 200 ft is no longer read with its 2000 to 10000 ft levels as the 200 to 1000 ft ones (16 to 30 dB low); a table without one of the distances, or with two columns matching one name, is refused.
Fixed: in the editor, after an example is saved elsewhere a change of absorption re-weights its run again instead of marking it out of date, and the sound playing is no longer said to be from before the latest changes.
Fixed: the editor’s Simulation page offers Parabolic equation and Frequencies behind barriers when only a long-term weather class brings the wind or the temperature gradient.
Changed: the editor’s Messages shows every error and the line that ends a run whichever Item is chosen.
Fixed: the tip of Frequencies behind barriers leaves the errors and the run times to the documentation (it still gave 12, 7.7 and 4.5 times a straight-leg map’s time, which are 8.2, 5.6 and 4.0 times since 4.2.0), and
fetch.extract_tarraisesOSError, nothttp.client.IncompleteRead, for a tar archive sent in chunks and cut short.Fixed: a walk with Doppler follows each path’s 1/distance from sample to sample, not every 20 ms: a source passing 1 m from a listener at 70 m/s rippled by up to 0.73 dB and is now within 0.01 dB of the exact field (1 kHz, 5 ms windows). The filters from place to place are joined linearly (triangles, not squared sines), which also moves a source heard through turbulence by up to 0.03 dB in its 5 ms levels.
Fixed: a demo’s tour fits each crossfade to the stops it joins; a stop shorter than the crossfade was played twice (+6 dB) or stopped the demo with an error.
Docs: a moving monopole’s convective amplification, which the engine leaves out, is 40 lg 1/(1 − M cos θ) dB (1.6 dB ahead at 30 m/s, 4.0 dB at 70 m/s), twice what the docs said.
Fixed: beside a bend of a thin screen’s top (a barrier’s path that turns by less than 15°) the level stepped by up to 5.4 dB, or 16 dB with the bend towards the source, and past a screen’s corner that is not square by up to 3.0 dB: each of an edge’s terms now takes its own part of the edge as BTM’s integrand weights it (within 0.04 dB of BTM beside bends, 0.17 dB past corners).
Fixed: a path that touches the very end of a fading edge, as with the source, the listener and a barrier’s top at one height, is no longer dropped (up to 5.6 dB low at 63 Hz).
Fixed: in wind or a temperature gradient the late rays counted again the curved paths off the ground that the early part already holds (those with more bounces than
max_order, up to ten), so over a ground that scatters little the level rose with the ray count: downwind of a low source, up to 8 dB over the fast field program at 500 Hz and 29 dB in the 250 Hz ground dip with 50,000 rays. They now leave out what the early part holds (within 0.05 dB of the program in those bands); still air and the shipped examples do not change.Fixed: with vector scattering (the default) a reflection’s specular share lost what plants absorb on its way but kept what they scatter, so behind 6.3 m of foliage scattering 0.3 per metre it came out 8.2 dB too loud. It now loses both (CPU, GPU and coverage maps); in
tree_beltone band of the open meadow drops by 0.2 to 0.27 dB in four of six seeds, the A-weighted levels by at most 0.03 dB.Fixed: the distance that sets how much a small or distant surface scatters (
scattering.from_size) restarted at each opening or partition a late ray, the listeners’ gather or a sound particle passed; it now runs from the source or the last reflection, as in the early part. An open face 1 m over a floor had taken 6.3 dB from the floor’s scattered sound at 125 Hz; inapartmentthe bed’s level drops by 0.04 dB(A).Fixed (results change): in wind or a temperature gradient a listener’s or map point’s bands up to 2 kHz depended on the others in its direction, whose ranges broke up the steps of the parabolic equation’s march and set its absorbing layer: by 0.12 dB at 63 Hz behind one listener 10 m nearer and 0.65 dB among 201 listeners 10 cm apart (a source 7 m and listeners 1.5 m up over a ground of 200 kPa s/m², 40 to 60 m downwind in
windy). Over the open ground each is now reached by a last step of its own off the march, under the absorbing layer of a run for it alone, and reads the same alone as among others. With that every level from the equation there moves once: by 0.07 dB or less down to 45 dB below free field in the checks, by up to 0.3 dB down to 60 dB below it and by more deeper in a shadow. Behind barriers and buildings the others still set the layer (see Known issues). A line of listeners about a metre apart or closer takes about 1.5 to 1.8 times as long as before the fix and a coverage map as long; listeners whose layers need different grids of heights no longer share a run (validation/long_term.py, five 50 to 800 m away, takes 1.3 times as long, three 10 to 20 m up 30 to 60 m from the source 1.5 times), and a listener alone that takes the wide-angle equation up to 1.3 times.Fixed (results change): a listener that the reflection off the ground reaches more than 15 degrees up takes the wide-angle parabolic equation wherever the parabolic equation runs, as behind barriers; it went by whether the air bent sound on its run’s grid, which changes with the band and with the run’s other listeners. In still air before a weather zone, 8 m from a source 2 m up, the early sound’s octave bands move by up to 0.11 dB and its third octaves by up to 0.28 dB (a one-off check).
Fixed: in wind, a listener or map point exactly in line with the source and a barrier’s end takes the weather’s effect on the paths round the end, whichever way the line rounds; it read up to 9.3 dB low downwind from 63 to 500 Hz and up to 12.3 dB low upwind from 63 Hz to 1 kHz.
Fixed: in wind, where the paths over edges keep their straight legs (a source or a listener on a building, a model file’s object in the way), the bent direct sound and ground reflection count only where their straight paths do; strips 0.3 to 3.5 m wide at the edges’ shadow boundaries read 4 to 10 dB off in every band.
Docs: three numbers that 4.2.0’s fix of the vector reflection’s specular share moved are given as the code has them:
apartment’s bed with the door open reads 62 dB(A), not 63, and the CPU’s next seed in validation.md differs by 1.32 dB and 20.6 ms intree_beltand by 2.04 dB and 1.9 ms on the hillside, a new draw of the rays’ noise.Fixed (results change): sound over two edges along a face that absorbs (
second_order, over a roof or a thick wall) was nearly cancelled in the bands that add in pressure, by the plane wave’s R = -1 at grazing incidence; the path now takes the face as rigid and the ground wave’s loss along it (Weyl-van der Pol), in every band. Behind a 6 m building absorbing 0.02 the level rises by 18 to 20 dB, to 0.1 to 0.8 dB below a rigid one’s, and the archway example’s first arrival behind the wall by 11.0 and 11.7 dB at 63 and 125 Hz; with faces absorbing 0.5 it changes by +11.0 dB at 63 Hz to -8.7 dB at 8 kHz.Fixed: a curved path that turns in the air just past a node of the sampled profile near the source’s or the listener’s height took that profile’s slope there, up to 15 dB off the smooth one’s; it now takes the fitted one. Against the smooth profiles (
validation/curved_rays.py, paths near a caustic apart) the most a path is off goes from 15.3 to 2.4 dB where the sound bends down and from 3.1 to 1.1 dB where it bends up; at 4 and 8 kHz a listener 300 m downwind of theopen_air_stagesource and 4.61 m up gains about 2 dB, and the shipped examples’ own listeners move by 0.02 dB at most.Fixed:
early.method: image_sourceswithmax_order3 or more (the default order) stopped with “need at least one array to concatenate” where the ground is the only reflecting plane, as in the open-air examples; the orders past the first now find no path and the levels are those ofmax_order: 1.Fixed:
write_ascwrote the grid’s origin, cell size and heights to 6 significant digits, moving a grid in UTM coordinates by up to 5 m (the northing rounded to 10 m) and heights of 1 to 10 km by up to 5 mm; it now writes each in the shortest form that reads back as the same number.Fixed: behind a barrier on the ground, where the reflection off the ground or a wall lies near the barrier’s foot, the edge fade gave the leg past the barrier terms of the leg it stops and kept paths over an edge that went on through a wall or another barrier: the level read up to 6.0 dB below the exact solution (7.1 dB with
both_sides) and stepped by 5.3 dB at 125 Hz over 2 cm, by up to 15.7 dB at 63 Hz behind two barriers. Behind one barrier the fade now adds nothing (withboth_sidesa straight one is within 0.2 dB of the exact solution from 63 Hz to 1 kHz); of the examples onlyrailway’s line and thestreet’s garden move, by up to 0.14 dB in a band.Fixed: a listener or map point exactly in line with the source and a building’s wall (or a block’s top), past the wall, heard the paths round both of its corners, one ray counted twice: 3 to 9.5 dB louder than 10 µm into the shadow, over one edge or two, with reflections up to the third order. It now has the paths and the level it has just beside the line in the shadow, in still air and in wind; a point on that wall itself keeps the level it had.
Fixed (results change): round an open model the wave solver’s matched layer lay a few cells past the model, the source and the listeners and sent back much of the sound that ran along it: over a rigid plate, with 40 cm cells, the 63 Hz band read 3.1 and 9.8 dB below the exact solution 80 and 160 m away. The grid now reaches past the source and the listeners on each side (not under an outdoor ground) by 15 % of how far the listeners lie from the source along that side, which brings those checks within 0.03 dB of it and costs cells wherever an open side of the model reaches less far (1.13 times as many on validation/wave_outdoor.py’s 63 Hz grids); a side that the model’s own surfaces close gets none.
Fixed (results change): the wave solver took its source and each point a listener samples from the eight cells round them, weighted trilinearly, which read the 125 Hz band up to 0.53 dB low by their place within a cell (8 points per wavelength at the band’s top). They now take a cubic through four cells a side (along a floor, a wall or flat ground, across it the two cells round them; next to other surfaces and within three cells of the source the eight cells as before), within 0.02 dB of a point source’s level from eight cells on in closed rooms, where the source’s signal also undoes the 19-point grid’s own gain on a point source (0.06 and 0.35 dB in the two bands), which that loss had hidden.
Changed:
validation/wave_solver.pyfades the cuts of its duct and open-model checks, which an abrupt cut had filled with sound from far above the bands they look at (in the duct, its first symmetric cross mode): the duct’s absorption is within 0.0002 of the grid’s theory (it showed 0.005 to 0.007), and the default matched layer returns 72 to 74 dB less than the sound (67 to 74). Its near-source table takes 50 directions and gives the gradient’s direction.Changed:
validation/wave_outdoor.py’s ground reaches 63 m (63 Hz band) and 60 m (125 Hz band) past the farthest listener instead of 30 m, so that the sound off its far end comes after the end of the run. Without that sound the 63 Hz band reads 0.14 to 0.29 dB lower at 200 m in every weather, and the 125 Hz band between 0.17 dB lower and 0.11 dB higher at 100 m (see Validation).Fixed (results change): with
simulation.wave.wind: moving, the wind blew through the model’s solids, which made the sound in a nearly closed space such as a room with an opening grow without bound; it now flows round them as an ideal fluid’s, so results near buildings change (the validation’s façade echo comes 0.41 ms sooner) and runs there take longer. Where weather zones change the wind along its way, it now passes from one zone’s wind to the next’s gradually, no cell gaining or losing air, which changes results there too. Each run in moving air is also watched and stops with an error should its sound grow, and a wave solver’s run that does not stay finite is refused.Fixed (results change): in the wave solver’s moving air, sound could build up in fast winds where the wind crosses the border of the grid’s absorbing layer, whose air is still (in the runs made, the slowest wind in which it did was 53 m/s, at 45 degrees to the grid over open ground on a grid 4 m tall). Each cell’s own sound now goes out through that border and silence comes in, which adds no energy there, and every result in moving air changes a little (the validation’s levels by up to 0.012 dB, its façade echo by 0.011 ms); as before, the border sends back about half the wind’s Mach number of the sound that leaves, 32 to 40 dB below it at 10 m/s (see Validation).
Fixed (results change): outdoors, the wave solver took the wind into a closed building round a shut-in source, whose air is still. In a 15 m/s wind a 6 m hut’s band energies were up to 1.7 dB off the still air’s with the effective speed of sound and up to 1.2 dB with
wind: moving, which also grew there instead of dying away. A closed model now takes still air, and the run says so (wave_still_inside).Fixed: the wave solver’s moving air let a closed rigid room’s mean pressure grow 2 to 2.6 times per 10,000 steps, from single-precision rounding in its carrying step, which now keeps a uniform pressure to the bit. Open models change by rounding only.
Fixed (results change): the wave solver stopped at the response’s end, where the bands’ filters still read the run that follows: the last 50 ms of the wave bands’ energy were up to 7.5 dB off (15.7 dB with third-octave bands) in a room of 2 % absorption. It now steps 0.2 s further (0.4 s with third-octave bands), which also keeps the 19-point stencil’s slow drift of a closed room’s mean pressure out of the bands.
Docs: the distances at which the wave solver’s gradient (X, Y, Z) near its source is further off a point source’s than read from eight cells, from a finer scan, and why: the source’s spread over four cells a side from 2 to 3.4 cells away, the listener’s reading beyond (Free field under Validation).
Docs: the cost of the wave solver’s moving air is given as four to five times the effective speed of sound’s over open ground and more round buildings (
validation/wave_wind.pymeasures 3.9 to 4.9 times, and 12 times round its façade’s building), where the help, the editor’s tip and the docs said three to four times.New validation:
validation/wave_walls.py, what the wave solver’s walls on cell faces cost: in seven placements of a rigid box moved within a cell or turned against the grid, its modes are up to 11 %, 6.4 % and 2.8 % off on the 63, 125 and 250 Hz bands’ grids (19-point stencil), and the level in front of a wall 0.4 cell off a cell face is 1.5 dB rms off at 8 to 16 cells per wavelength (see Validation).Changed: CI runs again by itself, up to twice, the jobs of a run that failed only because GitHub lost their runner or a step that fetches system packages failed (
.github/workflows/rerun.yml,tools/ci_rerun.py). A pull request runs one job on the Intel Mac, the tests with Python 3.10; main and tags run all four.Fixed: in wind or a temperature gradient the sound over the open ground at 1 and 2 kHz (630 Hz to 2.5 kHz in third octaves) came from rays outside the shadow, which were up to 3.2 dB off a fast field program in
windyand 8.5 dB inclear_nightand stepped by up to 14 dB a centimetre at a caustic; it now comes from the parabolic equation there too, at 24 frequencies an octave, which moves the examples’ bands by up to 2.8 dB (3.3 dB read from a third-octave run) and makes runs over open ground in wind take longer,open_air_stageabout 2.3 times as long on one core (see Outdoor scenes).Added:
simulation.early.open_ground_frequencies(Frequencies over open ground in the editor) picks how those 1 and 2 kHz bands are found: 24 frequencies an octave (the default, up to 0.75 dB off the program), 8 (4.1 dB, 2 to 3 times faster on one core) or the rays short of half the way to the shadow, as 4.2.1 did in octave bands (8.4 dB, 2.3 to 30 times faster); in the shadow 4 and 8 kHz, which deepen from those bands, move with them (by up to 0.33 dB within 60 dB of free field).helicopter_flyover, whose sound comes down steeply, sets the rays, there within 0.11 dB of the default.Fixed: a walk with Doppler takes each path’s filter (a microphone’s pattern, the ears, panning, the source’s directivity, the ground) at every place the walker passes, not every 20 ms alone: through a cardioid microphone a source passing 1 m away at 70 m/s missed the exact field’s level by 0.25 dB in 5 ms windows and is within 0.03 dB. The early part takes 1.1 to 1.2 times as long to render with places 1 m apart at 30 to 80 m/s (1.3 times in binaural), 2.6 times at 70 m/s with places 0.25 m apart (the whole rendering 1.4 times, 1.9 times in binaural).
Fixed: in binaural, a walking source’s sound reaches each ear that ear’s onset after it passes the head’s centre; the onset was added to the path’s delay, which shortened the time between the ears as the pitch rose. A source passing 2 m away at 70 m/s was 3.1 dB off the exact field’s level at 1 kHz in 5 ms windows and is within 0.06 dB.
Fixed: in wind, a source or a listener on a building’s roof left the building out of the parabolic equation, so past the roof’s edge the level stepped by up to 1.75 dB at 500 Hz in 10 cm and was up to 1.95 dB off a full-wave solution 200 m away. The equation now holds the building, as behind barriers (and the paths off roofs take the effect behind them, beside a building up to 3.7 dB off at 8 kHz before): within 0.12 dB of the full wave 27 to 43 m away, with a 0.12 dB step left where the equation takes over (see Known issues); such a listener takes 4 to 180 times as long, a map 35 times.
Fixed: across the shadow boundary of a barrier’s end or a building’s corner in wind the level stepped by up to 7.7 dB 60 m downwind and 18.8 dB 120 m downwind (1 mm either side); the paths round the end now change as the open ground’s field along their way does, and the step is at most 0.07 dB downwind and 0.22 dB 120 m upwind (where still air steps by 0.17 dB). Steps remain at 4 and 8 kHz with eddies (up to 1.08 dB) and in weather zones (0.76 dB), see Known issues; wind behind a building costs a run of the open ground’s equation a way (a coverage map round one, 32 % more CPU).
Fixed: in wind, the paths round a barrier’s end to a listener on a building’s roof (or from a source there) took the open ground’s field and the paths over its top past the end kept straight legs; they now take the runs the lit side’s paths take, so that 1.5 m over a roof 45 m from a source 1 m up (and the two swapped), 1 mm either side of the end’s shadow boundary, the weather’s effect steps by up to 0.57 dB in octave bands (1.28 dB in third octaves) instead of 6.4 dB, the rest from the paths over the roof’s edge, which only the lit side has (see Known issues).
Fixed: a leg along the ground, a rounding error above it, had no refraction shadow.
Fixed (results change): outdoors in wind and temperature gradients, out of the shadow, the parabolic equation’s bands carry its field at every frequency its runs took, on through the crossover above the top one and down to an octave below the lowest, so a ground dip at a join between bands is no longer missed (case 2 of the 1995 benchmark, worst from 2 to 3.45 kHz against the fast field program: 2.0 and 5.3 dB off 100 and 300 m downwind before, 0.19 and 0.45 dB now). The equation takes 11 to 79 % longer for a listener; C50 and C80 below 1 kHz move with it, C80 at 500 Hz away from a dense run of the equation, and the top crossover moves where the rays’ band beside it is off (see Known issues).
Fixed (results change): with eddies, the parabolic equation’s bands over the open ground took the rays’ coherence as one mean a band; they take it at each frequency the equation ran at, as the rays do, which brings 2 kHz in a breath of wind from 0.41 to 0.07 dB off still air’s image sources (0.12 dB without eddies). Downwind in
windy, a source 5 m and a listener 1.5 m up over grass 25 to 600 m away, the 250 Hz to 2 kHz bands move by up to 0.59 dB with thestrongeddies and 0.50 dB with themoderate.Fixed (results change): in wind, the paths round a barrier’s end or a building’s corner carry their way’s field through the crossovers between the parabolic equation’s bands, as a listener’s own field does: 1 mm either side of an end’s shadow boundary 60 and 120 m downwind the rendered third octaves from 1.25 to 3.15 kHz stepped by up to 1.0 dB more than in still air and now by at most 0.01 dB more. At listeners and coverage map points with such paths, what the band filters read up to 500 Hz moves with them: the strength by up to 0.05 dB, and C50 and C80 by up to 11 and 13 dB where they are 65 to 90 dB.
Known issues, measured: the steps at caustics in wind and temperature gradients against a fast field program on thirteen walks (up to 15.5 dB within a centimetre at 4 kHz), and a wall’s reflection there, whose paths are eigenrays in every band (see Known issues).
4.2.1¶
A user manual leads through the documentation from a first simulation to the advanced workflows, with an example scene to try at each step; the documentation is online at https://naelolaiz.github.io/acoustic-spaces/, with the manual as one PDF there and on every release.
Changed:
simulation.uncertainty.presets: iso_12999_2takes ISO 12999-2’s octave-band values (Table 2) from 250 Hz to 4 kHz instead of the mean of its third octaves’, so hard surfaces spread about twice as far there and strong absorbers a little less.Fixed: Getting started lists every install extra and what it adds in one table, which the README and the manual link to.
4.2.0¶
Error estimates: a material’s absorption uncertainty, a default one for the built-in materials and the rays’ random error, combined by Monte Carlo (GUM Supplement 1) into each room parameter’s uncertainty with its budget, beside the method’s errors over edges; barriers and buildings in wind or a temperature gradient through the parabolic equation; paths over a thin screen’s edge fading out past its free ends; the late sound scattered off the ground gathered from each listener; aircraft levels with Doc 29’s slow time weighting; the demos’ meters by ITU-R BS.1770-4 and their Ambisonics and loudspeaker tours decoded to the ears; walks crossfaded band by band; the editor’s Run demo window keeping its options, and the views following a run; and fixes to curved rays, the terrain’s mean plane, the wave solver, early reflections between rooms, coverage maps and sweep measurements.
Results change from 4.1.0: outdoors in wind or a temperature gradient behind and beside
barriers and buildings, past a thin screen’s free ends, near where a barrier cuts off a
reflection off soft ground (0.7 dB at most behind the railway example’s barrier), at
listeners whose ground reflection comes in more than 15° up, and where a curved ray
runs level near the
listener, the source or the ground (20 dB and more at worst); over uneven terrain, by up
to 2.1 dB at the hillside example’s listener; near where two grounds meet; beside
buildings, by up to 1.6 dB behind the railway example’s barrier; across weather zones;
and the late sound scattered off the ground near a listener, by up to 6.1
dB(A) at a place of the airport’s landing (BRAS’s CR1, CR2 and CR4 move by 0.04 JND at
most). With vector-based scattering (the default) the late part changes where a
reflection’s specular share could not go, by over 20 dB at 250 Hz behind the railway’s
barrier, and a little everywhere, maps as listeners: seven outdoor examples by 0.32 dB,
BRAS’s rooms by 0.3 JND and the indoor examples by 0.9 JND at most (mostly G down and
C80 up by 0.1 to 0.4 dB).
Indoors, early levels change where a reflection lands on the foot of a wall,
behind a wall between two rooms drawn as two faces (9 to 18 dB at 1 kHz behind a 20 dB
wall) and, with edge diffraction, near openings; the wave solver’s results within a few
metres of the source and next to thin surfaces; and coverage maps next to walls between
rooms, at their edges and, with the wave solver on, at 63 and 125 Hz. Stage support is
given only 0.9 to 1.1 m from a source, and a band whose G is undefined has a level. Doc
29’s aircraft are louder (JETW by 0.29 dB), an emission’s sound power is in IEC octaves
(the helicopter’s 63 Hz band 130.7 dB, was 134.8), the freight train has tank wagons and
the railway’s line source Schall 03’s directivity. The demos’ stored levels of
two-channel mixes rise by 3.0 to about 6.4 dB with the BS.1770 meters, and captions and
tour gains move by up to about 3.5 dB; walks without Doppler crossfade band by band, and
a walking source’s late part follows its early paths. deconvolve and measure cut
responses at other lengths. Scene files of 4.1.0 load unchanged; a scene with every
source or every listener left out is now refused by validate, inspect, pack and
the editor’s Run.
Error estimates
Added: a material’s
uncertainty(a standard uncertainty or a tolerance of its absorption), andmaterial_spread, the Monte Carlo spread (GUM Supplement 1) of each listener’s room parameters that it leads to, by re-weighting the traced run;simulateprints it and writes it intoparameters.json(--spread-draws, 400 by default, 0 for none), and the editor’s results give it in their tips.Added: the rays’ random error joins the absorption’s in the same Monte Carlo, as Student’s t from the late batches’ standard error, for a combined standard uncertainty and coverage interval with the budget of each input (
uncertaintyinparameters.json).Added: built-in materials, whose database does not name its sources, take an absorption uncertainty when they give none (
simulation.uncertainty.presets, Built-in materials in the editor): plus or minus 0.1 by default, or twice the reproducibility ISO 12999-2 gives for ISO 354, so a scene of presets gets a spread too. The editor shows a run’s results before it draws the spread.The spread takes the admittance of the paths that add in pressure between fits at nine absorptions of each material: fitting it for every draw took about 3 s a draw (20 minutes for 400 in the shoebox example).
Added: the spread lists the method’s errors over edges that it does not count, where a response has paths over edges, each with the validation’s largest error in each band.
Outdoors: barriers, wind, terrain and ground
Behind barriers and buildings in wind or a temperature gradient the paths in the vertical plane through the source and the listener take the weather’s effect from the parabolic equation run in every band with the barriers and buildings inside, and those round a barrier’s end or a building’s corner change as the open ground’s field along their own way does, for listeners in the open too (by up to 4 dB 3 to 5 m beside them, 17 dB a centimetre beside a building’s wall, where straight legs left a dip); they kept straight legs, so a barrier screened as much downwind as in still air. From 63 to 500 Hz the effect is within 0.71 dB of a full-wave solution behind one barrier or building in 5 m/s of wind and 0.74 dB downwind in 10 m/s, but up to 1.0 dB off behind a top 45 degrees up, 1.64 dB low upwind in 10 m/s and 1.73 and 4.95 dB off behind two barriers 10 and 40 m apart (
validation/barrier_wind.py); coverage maps behind barriers in wind take 8 to 11 times as long;simulation.early.barrier_frequencies(8, 16 or 32, the default; Frequencies behind barriers in the editor) takes fewer frequencies an octave at 4 and 8 kHz behind barriers, for less time and a weather’s effect further off there, by as much asdocs/outdoor.mdgives for each.A path over the edge of a thin screen now fades out past a free end of the edge, by the part of its Fresnel zone the edge covers as in BTM, instead of stopping there: past a screen’s corner the level is within 0.1 dB of the exact solution (it stepped by 2.8 dB), and 10 cm across a barrier end’s shadow boundary it steps by 0.42 dB in still air (was 4.5 dB) and, in wind, by 1.0 dB from 63 to 500 Hz (was 5.0 dB) and 0.91 dB 120 m upwind (was 27 dB). A solid block’s edges still stop at its corners.
With the default narrow-angle parabolic equation, outdoors in wind or temperature gradients, the listeners whose reflection off the ground comes in more than 15° up (near a high source or a high listener) get the wide-angle one’s level: against a fast field program, in 17 weathers, grounds and heights, their octave bands are 0.03 dB off on average and 0.10 dB at most (0.16 and 0.87 dB before, for a listener alone), their third octaves 0.04 and 0.32 dB (0.41 and 2.67 dB). No other listener’s level moves, and the steep listeners’ bands take about four times the work.
Fixed: a curved ray’s level near where it runs level at the listener, the source, the ground, a table’s row or the edge of a layer whose speed does not change with height came out several dB off, 20 dB and more at worst, and many such rays were missed; the parts of its spread that grow without bound there are now taken in closed form where that proves the steadier way, a ray whose spread no way gives steadily is left out, and the direct ray through air whose speed does not change with height is the straight one (a ray that turns once between ends at about the same height now follows a preset’s sampled profile, up to 0.33 dB off its smooth one). Neither a ground that absorbs nothing (in the parabolic equation) nor a layer whose speed changes by less than 5.6e-9 m/s per metre sets off a RuntimeWarning any more.
Fixed: beside a building, a reflection off the plane of its roof from below, out in the open air, took edge terms: its corners stood on that plane, and the roof’s edge explained a reflection off the back of the roof. Behind the railway example’s barrier the horn’s level was up to 1.6 dB too high (250 Hz) and stepped by up to 3.8 dB (8 kHz) where such a path met an end of the house’s wall.
Fixed: with weather zones, the shares of a path’s way under each weather, rounded one by one, could add up to 1 ± 0.005 (± 0.01 across five weathers), scaling the speed of sound in still air by as much; they are now divided by their sum. Over 300 m across three weathers, a third each, the direct sound arrives at 868.5 ms (was 864.2) and the levels at 1 to 8 kHz rise by 0.1 to 0.3 dB; the lakeside example is unchanged.
Fixed: over a terrain, the mean plane the early sound reflects off is the least-squares line of the whole ground profile, as CNOSSOS-EU fits it (ISO/TR 17534-4’s TC05 and TC06), not of the corners where the terrain’s triangles cut the profile, which moved it with the grid. Early levels in sight over uneven ground change, by up to 2.1 dB (1 kHz) at the hillside example’s listener on the slope.
Fixed: a ground reflection near where two grounds meet (a runway in a meadow) weighs each by the exact area of its Fresnel zone it covers, and a zone at the ground’s edge goes on past it; the shares were counted on 96 points, so a pass-by’s level stepped by up to 1.75 dB as the reflection crossed the runway’s edge. The airport’s departure heard at approach_2000m reads 0.2 dB(A) lower (L_ASmax 51.9).
Fixed: where a barrier’s top or end cuts off a reflection off a ground with an impedance, the reflection and its edge term now join: the term reflected off the ground as the wave from the edge does even on the boundary, where the reflection takes the whole path’s factor, so behind the railway example’s barrier the horn’s level stepped by up to 1.2 dB (250 Hz) as it passed the barrier’s end.
Fixed: one late ray scattered off the ground next to a listener could lift a place of an outdoor walk by up to 6.1 dB(A) (the airport’s landing heard at approach_2000m), and most runs found too little of that scattered sound far from the source. It is now shared between the rays and 16384 directions gathered from each listener, by how densely each finds it (
stochastic_energy(gather=...), each batch with its share): on that landing the late part lifts a place by 0.2 dB(A) at most, the aircraft validation’s highest unweighted level far out falls by up to 0.7 dB, and BRAS’s CR1, CR2 and CR4 move by 0.04 JND at most.Fixed: with vector-based scattering (the default) the detector spheres counted a reflection’s specular share on the ray’s next leg, which leans towards the wall’s normal, so that share also reached places no specular path goes: behind the railway example’s barrier one ray made the late part of the horn’s pass over 20 dB louder at 250 Hz, heard as a burst. The share now goes along the exact specular direction on a leg of its own up to the next face, for the listeners and the coverage maps alike; seven outdoor examples move by 0.32 dB at most (the mean over seven seeds), BRAS’s rooms by 0.3 JND at most over their pairs (C50 and D50 up, Ts down, typically by 0.1 JND), and seven indoor examples by 0.9 JND at most (the studio’s EDT at 63 Hz; G down by 0.4 dB at most, C80 up by 0.6 dB at most), where the results with and without the early stage’s reflections now agree better.
Maps over a long term gain the long-term insertion loss of what
insertion_loss_ofnames (long_term_insertion_loss, dB per band, andlong_term_insertion_loss_a, dB(A)), from the runs with and without it that every class already made. Their other values,insertion_lossamong them, are the scene’s own weather’s, and the levels in each class and period and relative to free field over the long term are at listeners only, as the docs now say.
Indoors: early reflections, the wave solver and coverage maps
Fixed: an early reflection whose point lies exactly on the foot of a wall, partition or barrier standing on the surface it reflects off (in a corner too, or where such walls meet), and which goes on behind it, now meets it there as it does just to one side of the point: through it with its transmission loss, or stopped by a rigid one; it passed whole (with image sources twice). In the apartment example the bedroom column of a 0.7 m map at x = 8.0 m read 51 dB(A) instead of about 45 like its neighbours, and behind a 3 m barrier a listener placed symmetrically to the source got 6 to 27 dB more early sound than 1 cm either side. Finding the early reflections takes 1.2 to 1.8 times as long in the examples; with edge diffraction the early part takes up to 3 % longer.
Fixed: where the wall between two rooms is drawn as two coincident faces, one for each room, no early path reflects off both faces at one point any more: from one room into the other that path was the direct sound at full level, without the wall’s transmission loss, and with image sources from the third order on it came back into the source’s room as the reflection off the wall once more. Behind such a 20 dB wall it had raised the early level at 1 kHz by 9 to 18 dB, and in front of it, depending on the listener, by about 0.04 to 0.25 dB up to the third order and 0.3 to 0.9 dB up to the fifth, more near the wall.
Fixed: no early path reflects off an opening any more; those paths carried no energy, but they took the place of other paths with the same image and, with edge diffraction, of the reflection bending round the opening’s edge. With edge diffraction, levels across a 10 m x 8 m room with a 2 m doorway move by as much as 2.4 dB at 63 Hz, 1.6 dB at 125 Hz and 0.2 dB above where the reflection off the doorway’s wall lands away from the doorway’s sides, by up to 4.3 dB at 63 Hz and 2.3 dB at 125 Hz where it lands on one, and no longer step by 2 dB where the reflection crosses a side.
Fixed: the wave solver took the pressure gradient across a thin surface with air behind it (a screen, a balcony’s front) at a listener within a grid cell of it, so the low end of Ambisonics, binaural and loudspeaker outputs and LF and LFC in the wave bands came out wrong there; it now takes it on the listener’s side. 0.1 m in front of a rigid screen the figure of eight facing it falls from 0.6 to 6.2 dB below the pressure in the 63 Hz band with the 7-point stencil (from 0.4 dB above to 14.7 dB below with the 19-point one), as in front of a solid block.
Fixed: the wave solver’s responses started at emission, though the source’s pulse is centred on it, so listeners within 16 time steps’ travel of the source (2.7 m on the default grid, 19-point up to 125 Hz, and 5.6 m on a 19-point grid up to 63 Hz) lost the pulse’s start; they now keep it. 0.3 to 1.5 m from the source the direct sound moves by -0.87 to +0.16 dB on the default grid and -0.41 to +0.97 dB on the 63 Hz one, and the figures of eight’s ratio to it by up to 1.9 and 2.1 dB; from 2 m on, both move by less than 0.01 dB on the default grid and up to 0.1 dB on the 63 Hz one.
Fixed: the wave solver’s source cut off at a fifth of the grid’s sample rate with either stencil, which took up to 1.6 dB off the top of the highest band with the 19-point stencil (whose time step is 1.5 times as long) at 5 cells per wavelength; the cutoff now scales with the time step (with the pulse’s start kept, that band rises there by up to 0.13 dB in octave bands and 0.49 dB in third-octave bands, and moves by 0.0033 dB at most at the default 8 cells per wavelength).
Fixed: the wave solver’s figures of eight (X, Y and Z: the low end’s direction in spatial outputs, and LF, LFC and LJ) came out 0.2 to 0.5 dB weak in the highest band; a fourth-order gradient and the velocity integral’s gain divided out bring them within 0.04 dB of a point source’s 4 m away along a grid axis or a diagonal, with either stencil. Within about 2.5 cells of the source they now depend more on where it lies in its cell; on the grid up to 125 Hz, 0.3 to 0.6 m away along a grid axis, they move by -8.5 to +1.5 dB in the 63 Hz band with the 19-point stencil and -3.1 to +1.4 dB with the 7-point one. With the two fixes above, LJ at seat-a in the shoebox example rises by 0.14 and 0.50 dB at 63 and 125 Hz (0.14 and 0.49 dB from this one alone).
Fixed: a coverage-map column that a wall between two rooms crosses had one point, whose slab counted the rays of both rooms; each room’s part of it that holds at least a fifth of it now gets a point that counts that room’s rays alone. Mapped 0.7 m apart, the apartment example’s bedroom points next to the partition read 65 dB(A); they now read 43 to 49 dB(A), 43.8 dB(A) where a listener simulated there gets 43.8 dB(A).
Fixed: a map slab’s volume counted the part of its column in the room in fifths, and all of it where the column reached into a building; it now counts the part in the air of its room, from 24 by 24 samples. Mapped 1.25 or 1.5 m apart, the shoebox example’s edge columns move by 0.2 to 0.5 dB(A); in the street example mapped 3 m apart, 12 columns of its canyon that reach into a building rise by 0.1 to 1.1 dB(A) for its traffic (up to 2 dB at 1 kHz).
Fixed: coverage maps of scenes with the wave solver on, which maps do not run, add the early paths in pressure up to
coherent_max_band_hzas documented, rather than in energy. In theroom_modesexample’s map G rises by 1.8 dB on average at 63 Hz and 1.2 dB at 125 Hz (up to 4.7 and 3.5 dB), C50 and C80 by up to 7.6 and 6.6 dB, and EDT falls by up to 31 %; the bands from 250 Hz up do not move.Stage support (ST early and late) is given only for listeners 0.9 to 1.1 m from a source, as ISO 3382-1 measures it 1 m away. Given up to 1.5 m at the listener’s own distance, it was off by up to 20 lg r (about +3 dB at 1.5 m, -5.5 dB at 0.5 m, 13 to 26 dB low at the singers’ ears and microphones); those listeners, the orchestra’s conductor, the stadium’s drummer and the church’s listener in the choir (for five of its seven singers) no longer get it.
A still listener’s level (
spl_dbandspl_a_dbinparameters.json, with the long-term levels, and the editor’s results) and the scene demo’s captions takepass_by’s rule (Response.levels_or_held): in a band whose G is undefined, the level of the energy the response holds. Levels that were empty there now have a value; only where no sound arrives they stay empty, and levels given before do not change.
Aircraft, helicopters, road vehicles and trains
Doc 29’s aircraft are louder in a scene, JETW by 0.29 dB: their NPD L_Amax is read with slow time weighting, so
doc29.source.sound_power_dbnow raises the level 1000 ft away by what that weighting takes off the NPD curves’ own pass (slow_drop_db). JETW’s sound power goes from 155.7 to 156.0 dB(A) departing and from 145.4 to 145.7 dB(A) arriving, and the airport example’s sources with it.Pass-by levels as Doc 29 measures them:
PassBy.slow_aandparameters.event_levelsgive L_ASmax (slow time weighting) and the SEL over the 10 dB-down time, which the flight demo’s captions andvalidation/aircraft_noise.pynow report. Against JETW’s NPD curves the engine’s L_ASmax is 2.6 to 2.8 dB under at 200 ft (its quasi-static L_Amax was 0.9 to 1.0 dB under) and within 0.2 dB at 1000 ft; the arrival’s SEL is 1.4 to 1.9 dB under (was 1.3 to 1.8), the departure’s still within 0.3 dB, and the airport’s flights move by up to 0.6 dB.Doc 29’s maximum level at a receiver alongside a segment (
doc29.segment_levels,doc29.event) takes the equivalent level path’s elevation angle, lateral displacement and depression angle, as the exposure does (Volume 2 section 4.5.5). L_Amax moves by up to 0.064 dB on Volume 3’s grids and receptors under the jets’ departures, 0.017 dB under PROP’s and 0.006 dB under the arrivals (SEL does not move), and the airport example’s departure at lateral_450m is 85.1 dB(A) by Doc 29 (was 85.0).doc29.eventwarns (doc29.NpdExtrapolationWarning) when a segment’s power lies so far outside the NPD curves’ that their extrapolated level departs more than 5 dB from the nearest curve’s (Volume 2 section 4.2); in Volume 3’s cases and the airport example’s flights it departs 2.3 dB at most.anp.Database.profilerefuses to fly procedural steps, which give thrust in lb, for an aircraft whose NPD curves are in another power parameter (Aircraft.power_in_lb).doc29.read_fixed_point_profilesreads a column’s unit from round or square brackets (m, ft, feet; m/s, kt, kts, knots) and stops on an unknown or missing one instead of taking metres or m/s.read_aircraftandnoise.installation_dbtake a lateral directivity by how it starts (“Wing-mounted” is a wing installation); any other is an error, where it gave no installation correction.performance.departurestops with an error on a climb step without the thrust to climb, with more thrust than weight or into a headwind that leaves it no climb angle (eq. B-21, B-22), instead of flying the profile backwards or failing inmath.asin.A walk whose first two places stand one above the other no longer flies the airport’s jets to NaN positions:
Track.throughleaves out a point within 1 µm of the one before it, andTrack.atandTrack.extendedpass over legs of no length.Volume 3’s impedance adjustment converts its 29.92 inHg with Appendix B2.2’s 3.38653131 kPa per inHg (
doc29.performance.KPA_PER_INHG), not 3.386389: its levels rise by 0.0002 dB, the 99 levels at its receptors now within 0.0055 dB of table B-1 (was 0.0057 dB).doc29.segment_levelsanddoc29.eventdocument theirimpedance, which the caller passes fromdoc29.impedance_dbfor the aerodrome; its default, 0, is the NPD curves’ reference day (25 °C), 0.074 dB under the standard atmosphere at sea level.Changed: an emission’s sound power and directivity (
Emission.sound_power_db,directivity) are now in IEC 61260-1 octaves between their band edges, 44.7 Hz to 11.2 kHz, not in the auralization’s crossover bands, whose 63 Hz band reached down to 0 Hz and whose 8 kHz band up to half the sample rate. The helicopter’slevel_dbandhelicopter.csvare regenerated: in helicopter_flyover 130.7 dB at 63 Hz (was 134.8, with the main rotor’s 19.7 and 39.4 Hz tones), 120.2 dB at 8 kHz (121.7) and 132.9 dB(A) (133.1); in stadium_concert 129.8 (134.0), 120.1 (121.6) and 132.8 dB(A) (132.9).emission.Mixturerefuses signals without samples and a sample rate that is not positive, which failed later with a division by zero.Fixed:
airfoil_noise.thicknessesgave a lightly tripped boundary layer (LIGHTLY_TRIPPED) the heavily tripped suction-side fits instead of the untripped ones of Brooks, Pope and Marcolini’s code: a 0.3048 m chord in 71.3 m/s was 1.3 dB too loud in TBL-TE noise at 8 degrees and 3.4 dB too quiet at 12. The helicopter’s blades, tripped or not, are unchanged.Fixed: the helicopter demo’s
flightlengthened the responses by the helicopter’s farthest distance, which the simulation already adds outdoors, so they ran 6.1 s, not 4.0 s; its walk at the mix position takes 121 s, not 174 s, at the same level (L_Amax 81.5 dB(A), 81.6 before).The pass-by demos’ freight train has 5 tank wagons among its 24, Schall 03’s share of 20 % when a freight train’s make-up is not known (4.2), their bodies ringing 4 m up: 116.4 dB(A) of sound power there instead of 111.8, 131.9 dB(A) in all instead of 131.8. Behind the railway’s barrier its SEL goes from 79.6 to 82.6 dB(A); at the other listeners its levels rise by 0.1 to 0.4 dB.
The railway example’s line source has Schall 03’s directivity round the track (equation 8,
line.csv,acoustic_spaces.demos.rail.directivity("left")) and so 0.28 dB more sound power, its points 2.5 m apart instead of 10 m: in a free field it gives Schall 03’s sum within 0.01 dB(A), not 0.2 to 0.4 dB(A) over it. Over the example’s ground it is 0.2 to 0.7 dB(A) louder, 1.0 dB(A) quieter behind the barrier, and the example takes 2.3 times as long to simulate.The pass-by captions average a horn’s sound power over 125 ms (the Fast time constant) before each place takes its share, so a short walk step no longer reads the beats between its notes as louder blasts (3 to 5 dB over at 5 cm). The street’s car with its horn now peaks at 97.0, 93.9 and 80.5 dB(A) (was 97.1, 94.0 and 81.6), the fire engine at 111.0 dB(A) on the balcony (was 111.1).
acoustic_spaces.vibrationraises a ValueError that says what it takes for a speed or a distance from plant that is not a finite number above 0, a land use or a kind of events its tables do not have, and special trackwork that is not a finite 0 m or more away (None for none near). A train’sUnitraises one for a tank wagon or air conditioning that its category’s data sheet does not have, instead of leaving it out.
Walks and demos
Walks with Doppler: a walking source’s late part is faded from place to place as its sound is sent, so that it follows the early paths from the same place, not from where the source had got to when it was heard (27 m on for the helicopter overhead). Noise through small cases moves by up to 0.8 dB in 125 ms windows (a 240 m flyover 150 m up, places 4 m apart: L_max +0.4 dB, energy +0.04 dB; the street’s car: 0.7, 0.01 and 0.04 dB), the validation’s pass-by over grass by up to 0.04 dB in a band; walking listeners are unchanged.
Walks without Doppler crossfade from one place to the next band by band, corrected by how alike the two places sound, so that each octave band’s power runs on (
acoustic_spaces.render.crossfaded, shared with the demos’ tours). Against the exact field of a pass-by, pure tones’ level misses by at most 2.7 to 4.1 dB instead of 4.2 to 7.0 dB, 0.2 to 2.4 dB on average instead of 1.3 to 2.7 dB, and for the train at 4 kHz by 12.2 instead of 10.9 dB, where the field’s two paths cancel for a moment.A walk as fast as sound or faster now stops with an error without Doppler too, as it did with Doppler: it played as places crossfaded one after another, as if shock waves did not exist.
Fixed: the demos’ meters (
level_db,loudness_db) add up a mix’s channel powers as ITU-R BS.1770-4 does (side surrounds weighted 1.41, the LFE left out) and read Ambisonics on W alone; they averaged the channels first. Stored levels of binaural and microphone-pair mixes rise by 3.0 to about 6.4 dB with how unlike the two channels are, so captions’ “dB(A) against the loudest” and tour gains move by up to about 3.5 dB; an order-1 AmbiX plane wave reads 6.0 dB higher (from behind it read silent), a 5.0 front speaker alone 14.0 dB higher, and one-channel mono mixes do not move.A tour turns a stop up by 40 dB at most, and a silent one not at all; a silent stop got the whole difference to the loudest, over 200 dB.
Fixed: a demo’s
tour.wavandtour.mp4of an Ambisonics or loudspeaker mix (--format ambix3,5.0…) are now decoded to the two ears (the binaural decoder, virtual speakers), through the scene’s HRTFs or the spherical head model; they played the first two channels, W and Y as left and right, and nothing of a 5.0 centre. Binaural and mono tours are unchanged.Fixed:
--formattakes an output’s label (ambix3,5.0) everywhere: the demos’ simulate step askssimulatefor it as--outputspells it (ambisonics:3,loudspeakers:5.0), and the demos that render the mixes themselves (the demo of any scene when something moves, the pass-by demos, the road works heard--indoors) take it too; both stopped on a raw validation error, and a format they do not know now says what to give.Fixed:
allin the other demos took mixes made through other HRTFs (--hrtf) as done, and without--hrtfa simulation through an HRTF file too, mixing through it; their videos now refuse mixes of another--formator HRTF. Mixes from before say neither, so they are made again.The demos’ JSON files (walk.json and the others) are strict JSON: a level where nothing is heard, as between a horn’s blasts, is written null instead of -Infinity, which browsers refuse.
--indoorstakes a room’s level outdoors 2 m in front of its wall, EN ISO 12354-3’s L_1,2m, simulating one more listener there where its listener stands nearer (the street example has it,balcony_2m, ticked off): the offices’ room behind the street’s balcony, 1 m out, moves by up to 2.1 dB at 63 to 250 Hz and 0.2 dB(A) for the tram and the road works. No documented number changes.The video’s messages say what it shows at each moment,
stage, car @ 3.40 s -> seat: encoding the video, 12.1 of 60.0 s, in place of ffmpeg’s frame counter.Fixed: the gallery walk demo stopped, after downloading the talk, when a talker’s take was shorter than the walk and the turn (F4’s, 36 s); the guide now says it again from its start, 2 s after it ends. Its walk and turn are made again for another
--hrtf, wherealltook them as done and the video stopped for want of their sound.Fixed: a still tour’s video is made in a work folder whose path holds an apostrophe, where ffmpeg stopped with exit status 254, and
demos.audio.envelopegives a signal shorter than its 0.5 s window its RMS (an empty one 0) instead of raising IndexError.Fixed: the lakeside example sets
simulation.seed, so its demo’s weather effect is the same every run; unseeded runs differed by up to 4.2 dB at 125 Hz at the bedroom window.Fixed: a demo given
--hrtfwith a file that is not there is refused before its first step; it failed only when a later step read the file.Fixed: the listening room’s
bass.pyreads the scene it is given, the open one from the editor’s Run demo window, instead of always the example’s.On Windows, a demo whose scene lies on another drive than its example no longer stops when it writes a siren’s, horn’s, train’s, helicopter’s or aircraft’s directivity table: the scene names the table in full.
Fixed: the pass-by graphs’ time labels are at least 60 px apart (every 10 s for the airport’s flights); they were every 2 s and overlapped.
Measuring
Fixed: with repeats,
deconvolveandmeasurecut a response asked for longer than the silence to the silence, as the next sweep’s harmonic distortion lands right after it; they cut it at the 2nd harmonic’s peak, which let the 3rd and higher (for a loudspeaker adding 0.2 x³ to a -6 dBFS sweep, 0.64 % of the linear peak) and part of the 2nd into the average.Fixed: with one sweep,
deconvolvecuts a response to the recording’s length less the sweep’s, past which the response lacks its high frequencies (measure: to the silence); the cut was the 2nd harmonic’s delay short of the period, whatever the recording’s length. A note says when a response comes out shorter than--length.Fixed: analysing an impulse response, a band whose lower edge is at or above half the sample rate (the 8 kHz band at sample rates below 11.25 kHz) has no parameters, nor has the STI or IACC there; its filter’s skirt over the band below gave it a T30, a G and the rest.
analyze,measure,compareandcalibratesay so, with the band that reaches past half the rate, which is analysed from its part below it.Fixed: a WAV file over 4 GiB (16 channels for 24 minutes at 48 kHz, say) is written as RF64 (EBU Tech 3306); it failed after the whole render. RF64 and BW64 files are read too.
Editor
The Run demo window keeps the options changed in it for the next time it opens, each demo its own (the listeners, the sources’ files, what moves and the format per scene file), names them above Run, and Defaults puts them back.
View > Follow the run rings the sources and listeners a run’s step works on, a demo’s too, and draws a walking source and the run’s other walkers where the place being simulated has them, on the walks the run simulates (
Demo.walks); while a demo’s video is encoded, what the video shows at that moment; a head a demo turns faces as it does then, its arrows and directivity balloon with it.Fixed: the editor’s material page lowers a transmission loss typed in, and an absorption the transmission gives follows it; it kept the higher one.
Fixed: another preset, file or entry, or a lower transmission loss, written for a material in the Scene YAML panel or in a scene file the editor saves now gives its coefficients and source; the old ones stayed, and the lower loss was refused as more than the absorption. Scene files and the panel write a material’s absorption, scattering and source only where its preset, file, entry or transmission does not give them, and a scattering entry by its name.
Fixed: Scene YAML text applied after the scene was saved in another folder names the files it named; it took files of the same name in the new folder, or was refused where there were none.
Fixed: an example saved elsewhere names the copies of its sources’ signals and its recording, as it does its other files; it named the originals.
Fixed: undo past a Save As that copied the scene’s files keeps the scene on the copies, instead of going back to the files it left (the package’s, for an example).
Fixed: an example’s demo that the editor’s Run demo… runs away from the example (in
~/acoustic-spaces/<example>, or beside a copy saved elsewhere) wrote its scenes, results and work among the example’s files, an installed copy’s too, and read the example’s scene rather than a copy saved asscene.yaml; it now gets every path in full, as the Command box shows (Demo.in_place).Fixed: the Run demo, particle export, buildings-from-a-map and Download HRTFs windows are deleted once answered, instead of staying behind until the editor closes.
Fixed: a demo’s messages pass between its process and the editor as UTF-8, so a folder name outside Windows’ code page no longer stops the demo.
Fixed: the 3D view’s names of walkers now move with them in the Movement panel’s preview; they stayed where each walk starts.
Scene files, the command line and downloads
A scene with every source, or every listener, left out (
enabled: false) is a problem:validate,inspect,packand the editor’s Run (F5) and Pack refuse it, assimulatedid; runs that name their sources or listeners, walks, which take their walker in any case, and maps, which need no listener, still take it.Fixed:
simulateandcoverage_mapcheck the sources and listeners they are given by name when those are left out (enabled: false) too: one outside the room ran.Fixed: a walking source is heard at a listener left out that the walk names (
walkthrough --listener), as other runs that name one; with every listener left out and none named, the error says so in place ofpick a listener:.Fixed:
simulate --withoutandmap --withoutare checked against the scene before anything runs, asinsertion_loss_ofis: an unknown name failed only after the first run, and a name given twice was taken.A coefficient CSV laid out by columns, or with more than three rows, is refused with the layout expected: the first loaded as absorption 1 in every band, the second without its rows past the third. Lines of separators alone (
,,,, a spreadsheet’s empty rows) are skipped.A scene’s
recordingwithout a suffix that is neither a free recording nor a file there (a free name mistyped) is refused, naming the free ones.A scene whose air pressure lies outside the range ISO 9613-1:1993 states the air absorption’s accuracy for (0.794 to 157.7 kPa in octave bands, 1 to 125.3 kPa in third-octave bands) is warned about.
Fixed: a simulation refused for its memory says how much it would need with each setting it names lowered on its own, the least first; it named them without how much each saves (more
late.batchesneed more memory, not less).Fixed:
packrefuses a file whose name holds:or\, which it packed but no computer could unpack, and two files, or a file and a folder, whose names macOS or Windows take for one (differing only in case, or in the way a letter is encoded); folders from outside the scene’s folder are named apart from the others regardless of case, and from files.pack --helpand the docs list the files sources play and the scene’s recording among the files a pack holds.Fixed: a download cut short (fewer bytes than the server announced) is refused and removed; it was kept as a whole HRTF set or recording, which later runs then used. One sent in chunks and cut short fails with an
OSErrortoo, as callers expect (hrtf downloadended with a traceback), and a remote archive’s block cut short is asked for again.Fixed: a tar archive cut short or damaged raises
OSError, as other failed downloads do, instead oftarfile.ReadError.
Speed and memory
The GPU traces a late part with face groups too, counting each ray’s reflections off each group: the spread from the absorption’s uncertainty, which
simulatedraws by default for a scene of built-in materials, the editor’s runs and material calibration no longer send the late part to the CPU wherelate.deviceasks for the GPU. Its moments agree with the CPU’s within the noise of the rays.Coverage maps and runs in wind take less time, with the same results bit for bit: the parabolic equation’s steps make fewer new arrays, and the eigenray searches it repeated are kept (on one core, a map of 96 points over grass in 8 m/s takes 14 % less time, behind two barriers 22 % less).
Walks with Doppler keep each place’s path filters only while the blocks that use them are rendered, and from three channels on as the paths’ shaping and the format’s gains: the street example’s car walk at the default 0.25 m step holds 0.74 GiB of them instead of up to 23.2 GiB in third-order Ambisonics and 15.9 GiB in 7.0.4, with the same results.
simulate_movementmerges a mover’s moments into places in milliseconds, where 4000 moments took 12 s.Walks with Doppler round each path’s reading cutoff down to 32 / n of the Nyquist frequency, n the taps it needs on each side, so that the paths share a few tables of taps: a 2 s street pass builds 8 instead of 188 and renders about a fifth faster. The sound changes almost only above 20 kHz (99 % of the difference lies above 20.6 kHz): by about 4 to 7 % of the amplitude of white noise through it, under 0.1 % below 20 kHz.
Models load faster: the search for faces to split at T-junctions is vectorised and gives the same faces (a 2 km terrain builds in 0.3 s instead of 5.4 s).
Outdoor models make the wave solver’s grid once per source instead of twice: a grid without the matched layer (with
scheme: autothe isotropic stencil’s), made only to find that the sky lets the sound out, is gone (the street example’s set-up at 63 Hz takes 2.4 s instead of 6.9 s on two threads), with the same results. The isotropic stencil is now refused outdoors before any grid is made, also for a source shut in a building, which used to run.The late part’s tracer holds one copy of its results fewer, with the same results: its threads’ copies, the first taking the others’ sums, so on one thread (as for copies over 128 MiB) it holds half as much.
Fixed: with
late.device: both, the CPU traced about 8 % of the rays whatever its speed, as the GPU took 4,096 rays at least of each call of the default 10 batches; each now takes no more than its share of a call.The GPU tracer keeps its buffer of rays in flight (up to 46 MiB for one band, 88 MiB for eight) through a late part, instead of making and clearing it for each of the part’s calls (160 per source with the defaults).
A loudspeaker layout with height finds the faces of its hull among the triples around each corner of scipy’s Qhull hull, with the same gains: 96 speakers in 0.12 s instead of 7.5 s.
The helicopter demos’
helicopterandwalksteps find the rotors’ tones and BPM’s bands (rotorcraft.bpm_powers) once, not for each length: about 7 s faster, 15 s with--broadband bpm.The pass-by demos fill a train’s or tram’s sound files a point at a time in 32 bits: the railway example’s freight train, 50 points at the rail top for 10.8 s at 48 kHz, takes 104 MB instead of 415 MB.
Validation and docs
BRAS’s coupled rooms with the door open by 30.4° join the validation: the conversion joins the door leaf, which ran into the wall in BRAS’s model, to the wall and floor. T20 of the chamber’s sound in the laboratory comes 11 to 12 % long from 500 Hz to 4 kHz.
validation/edge_diffraction.pychecks the level on and near an edge’s shadow boundary through the whole early stage against the exact solution: in the bands that add in energy (indoors from 250 Hz up) it is 1.73 dB low at 250 Hz and 0.38 dB at 8 kHz on the boundary of a right-angled corner, and 2.22 and 0.56 dB on that of a wedge open over 225°, now a known issue in the validation report.Fixed: the exact solutions in
validation/that the edge diffraction is checked against ran at 343.0 m/s, not at their scenes’ 343.2 m/s; the published differences move by up to 0.02 dB, and the terrain ridge’s reference, which had not converged, by up to 0.2 dB from 2 kHz up.Docs: G of responses written with
simulate --normalizetakes the reference energy 0.01 ×ir_gain² (fromparameters.json), and the JNDs of T20, T30 and C50 are ISO 3382-1’s of EDT and C80 taken by analogy; Barron’s total and the first batch of split rays are given as the code has them.Docs: each caustic a curved ray touches advances its phase by a quarter period, as the engine has it (a factor j, time going as e^(jωt)), not delays it; a test now holds the sign, another the shadow zone’s residue series to the fast field program.
Docs: in a belt under 20 m deep the engine’s dense foliage takes more than ISO 9613-2’s table at 63 and 125 Hz (in every band under 10 m) and up to 1.8 dB less at 4 and 8 kHz, not a little less throughout; stems 0.3 m across, 0.08 per m², take 0.002 dB per metre at 63 Hz (was 0.003).
Docs: the stadium, helicopter and lakeside demos’ numbers and pictures come from new runs with the BS.1770 meters: on stage it is 6 to 18 dB(A) louder than in the audience (7 to 20 before), and the lake’s weather lifts the bedroom window 5.3 to 6.0 dB over three runs (7.1 before).
The road docs say that CNOSSOS-EU’s correction for road gradients is left out, that the sirens’ levels 7 m in front of a vehicle are taken 7 m from the siren, and that recordings are levelled unweighted; a test holds every coefficient of Table F-1.
Docs: JETW’s arrival NPD curves give SEL - L_Amax within 0.22 dB of a point source’s (not 0.1 dB), Volume 3 Part 2 has no arrivals of PROP, and the NPD curves’ microphone height is cited from Volume 2 sections 2.5 and 4.5.4.
Docs: the TSI table gives each unit’s whole pass over Tp, beside EN ISO 3095:2013’s own time windows for a coach or a wagon; the glazing’s traffic figure is rated from its octaves, 29 - 4 + 9.6 = 34.6 dB; the FTA curves’ docs say how
Curve.levelsreads them outside 10 to 300 ft.Docs: the AS350 B3’s main rotor turns at 394 rpm, the most in powered flight with its Arriel 2B engine and so the speed appendix J flies; its tail rotor’s 2,086 rpm is NASA’s, no longer assumed; its tones agree with Farassat’s formulation to 7.5 x 10⁻⁶ (main rotor) and 1.6 x 10⁻⁵ (tail rotor), not 10⁻⁵.
Fixed: the Viking HRTF set is credited to Spagnol, Miccini and Unnthórsson of Aalborg University and the University of Iceland, and described as a KEMAR dummy head with one of 20 pairs of replica pinnae, or none; it read Reykjavík University and 20 people.
4.1.0¶
Helicopters flying over, with the sound a source sends each way carried down each early path of a walk with Doppler; aircraft taking off and landing by ECAC Doc 29 in the new airport example; trucks, buses, motorcycles, a police car, a fire engine, trains and a tram in the pass-by demos, with horns on the car and the train, and road works in the street; the ground-borne vibration of trains, trams and road works by the FTA’s method, and the trains, the tram and the road works heard indoors; listeners as microphones; a standard error for every parameter; a stereo recording for a scene’s sound system, so that every example with a stage, a sound system or a talker plays out of the box; the listeners a demo plays picked for a run; the editor’s Scene YAML panel; and four of BRAS’s rooms in the validation report.
Results change from 4.0.1: EDT, in 99 of 152 octave-band EDTs of six examples, by up to
9.6 %; the late part by rounding, as its rays are traced in 10 batches; random
scattering where the bands’ coefficients differ; indoors, rendered responses and the
echogram’s 63 and 125 Hz bands where paths over edges arrive together; the treble of
binaural walks through measured ears and of listening in turbulence; the car’s sound
power, 2.1 to 2.7 dB(A) more; the tree belt and lakeside concert examples, traced to the
second order; the demos’ stereo sources and the lakeside concert’s subwoofers, by up to
3 dB, and tours at their crossfades; and the examples’ talkers, at 60 dB(A) 1 m in front
instead of 94 dB of sound power. Scene files of 4.0.1 load unchanged; stadium_pa is
now part of stadium_concert, gallery_walk no longer names the gallery, and
acoustic_spaces.demos.car is now acoustic_spaces.demos.road.
Helicopters, and the sound a source sends each way
The sound a source sends each way (
acoustic_spaces.emission): along a walk with Doppler,render_walkandrender_movementtake anEmission, and each early path carries what is sent the way it leaves the source (Mixture,directivity.on_grid,sphere_mean).emission.Periodic, a sound that repeats, kept as one period in each direction, which may turn with a rotor and be sent from off the source’s place;emission.Sumadds emissions up; everyEmissiongives itsdirectivity().air.densitygives the density of dry air.Emission.sound_power_dbgives an emission’s sound power per band,emission.Fadedfades one in and out, anddirectivity.write_directivity_csvwrites a directivity as a CSV table a scene reads.Helicopters (
acoustic_spaces.rotorcraft):trimbalances a helicopter in level flight, andemissiongives the tones and broadband noise (Pegg’s method) its rotors send each way. Flying the AS350 B3’s noise certification, the model gives 88.0 dB(A), 3.4 dB over the certified 84.6 dB(A) (validation/rotorcraft.py).A helicopter’s broadband noise by either of two models (
broadband): Pegg’s method (the default) or Brooks, Pope and Marcolini’s airfoil self-noise summed over the blade sections ("bpm",acoustic_spaces.airfoil_noise). In the AS350 B3’s certification overflight they fall on either side of the certified 84.6 dB(A): 88.0 with Pegg’s, 80.1 with the blade sections’.A helicopter flying over: the
helicopter_flyoverexample, an AS350 B3 150 m up over a meadow, andflyover_demo.py, which plays its rotors’ sound down each early path with Doppler (demos.passby.Vehicle,demos.helicopter).A helicopter over the stadium concert: the
stadium_concertscene has the AS350 B3 crossing 150 m up (--height 330keeps it 300 m above the roof, as the EU’s rules of the air ask without leave), andhelicopter_demo.pyadds its pass to each listener’s mix of the concert, both at their simulated levels. An example may have more than one demo (demos.demos_of,Demo.builds_on).
Aircraft and ground vibration
Aircraft noise at receivers by ECAC Doc 29 (5th edition, 2026) in the new
acoustic_spaces.doc29: one departure’s or arrival’s SEL and L_Amax from the NPD curves along its flight path’s segments, with Volume 3’s three hypothetical aircraft (the ANP database’s own may not be passed on), checked by the newvalidation/doc29_reference.pyagainst all of Volume 3’s reference levels.Doc 29’s flight performance:
doc29.performancemakes an aircraft’s profile from its procedural steps by Volume 2 Appendix B, matching all 26 of Volume 3 Part 2’s reference profiles, anddoc29.anp.read_databasereads a copy of the ANP database one downloads under its terms.Aircraft taking off and landing in a scene: the new
airportexample’sflight_demo.pyflies Doc 29’s hypothetical twin jet JETW past its listeners by Appendix B, its sound power and directivity from its NPD curves following its thrust (doc29.source,doc29.atmosphere). The newvalidation/aircraft_noise.pychecks the source against its NPD curves and the flights against Doc 29’s own levels; the flights agree within 0.9 dB in the air but for the arrival’s SEL (1.7 dB under), and part by up to 11.4 dB on the runway, where the engine reckons the ground in still air.Ground-borne vibration of trains, trams and construction plant, and the rumble it makes indoors, by the general method of the US FTA’s manual (2018) in the new
acoustic_spaces.vibration; the railway’s and the street’s newvibration.pyprint it at their buildings against the manual’s limits, checked by the newvalidation/fta_vibration.pyagainst the manual’s figure and worked example.
Microphones
Listeners can be microphones: a listener’s
patternis what its mono output picks up through, a first-order pattern or an XY, Blumlein or MS pair in two channels (spatial.Microphone), with a Microphone choice in the editor. The parameters and the other outputs stay an omnidirectional listener’s.A listener’s microphone can have a pattern that changes with frequency, one per octave or third-octave band, for a single capsule, an XY pair’s
capsuleor an MS pair’smid; the editor shows it as per band: ….sceneexportsOrientation,XYPair,BlumleinPairandMSPair.
Simulation
Standard errors: the late rays are traced in batches (
simulation.late.batches, 10 by default), and a jackknife over them gives each parameter its standard error, inparameters.json,simulate’s output and the editor’s results. The batches make the late part 10 to 16 % slower;batches: 1traces the rays at once.Outdoors, the places of a walking source share the parabolic equation’s runs where they stand at one height over one ground without zones (
simulate(share_waves=...)): an aircraft’s 26 places take 31 s instead of 89 s, with the same early levels to 10⁻⁷ dB. Still scenes are unchanged, bit for bit. A place waiting for its run keeps only its bent paths.A moving source renders faster: the paths at a moment of the eddies are worked out once, their own phases’ angles once per place, and the HRTF onsets without the HRIRs.
Fixes
EDT’s line fit starts at the onset in every response; it left out the slot holding the direct sound in 99 of 152 octave-band EDTs of six examples, by up to 9.6 %.
Random scattering (
simulation.scattering.method: random) traces each band with rays of its own where the bands’ coefficients differ, as vector-based scattering does (by_band): with one ray for all bands, its T30 in the BRAS rooms had standard errors 15 to 180 times vector’s. It takes 3 to 4 times as long as before;by_band: falsekeeps one ray for all bands.Rendered responses give each path over an edge a phase of its own above the bands that add in pressure, so the pieces of a diffracted wave no longer count their interference twice (forty pieces within a millisecond rendered more than 20 times their energy at 250 Hz).
The echogram’s 63 and 125 Hz bands no longer take the interference of paths over edges a second time from the 250 Hz band. The parameters at 125 Hz move where many diffracted paths arrive together, as in BRAS’s rooms (with the fitted absorption, the hall’s EDT goes from 13 to 3.8 % short).
Behind a barrier on a hard ground with
first_order, the path over the top with the reflection off the ground behind it is kept, where the level stepped by 0.35 dB at 63 Hz; withboth_sidesit follows the exact level within 0.05 dB. No example changes; BRAS CR2’s early level moves by up to 0.8 dB at 63 Hz.Cleaning a model also splits a face where its edge runs past a corner at which three faces meet, so the BRAS auditorium (CR4) is now closed (8,656.6 m³); no example’s model changes.
walkthrough.pass_byno longer gives NaN for L_Amax, the SEL and L_Aeq when a place’s G is undefined in a band; every other place’s levels are unchanged.A walk with Doppler in binaural through measured ears (a SOFA file, such as
kemar) lost up to 6.8 dB at 8 to 16 kHz for 40 ms at a time between places, audible as a vehicle passes; each ear’s onset now goes into that ear’s delay.In turbulence, a walk with Doppler left out the delay the eddies give each path’s top band, as in the
helicopter_flyoverexample’s flyovers; that delay now goes into the path’s delay.Listening in turbulence (
auralize, the editor’s Listen, walks without Doppler) cancelled the treble between the eddies’ moments, by up to 8.9 dB at 8 to 16 kHz; each early path now plays through its own delay, going from one moment’s to the next’s.The pass-by demos’ captions, stills and printouts gave a walk’s speed such as 50.004 km/h; they now give it to a tenth of a km/h (50 km/h, 72.5 km/h).
A pass-by demo run at
--speedwith some of a vehicle’s sources left out (--police-car siren,--fire-engine siren) stopped at its walk, as the sources left out kept the scene’s speed; they now walk at the run’s.The Run demo window, switched to another demo, could come out too short for that demo’s options, which then had to be scrolled to.
The airport’s flight demo stopped on the departure (“the departing jet’s sound is shorter than its pass”), its sound a sample shorter than its walk’s rounded times.
A stereo recording’s left and right sources played 0.2 to 3 dB over their sound power, in any scene’s demo and the stadium’s; the two channels now share one gain set from both gated together, not from the mixdown. The lakeside concert’s subwoofers, fed the song’s mixdown, played up to 3 dB under theirs; they are now brought to the same level. Feeds made before in a work folder are made again.
A tour played up to 3 dB louder at a crossfade between two seats of one piece; the crossfade now keeps the power running on however alike the two mixes are.
Examples and demos
The pass-by demos drive more vehicles past (
passby_demo.py --vehicle): an articulated truck inroadside, and a city bus, a motorcycle and a police car with its siren on instreet, each with CNOSSOS-EU’s sound power of its category (acoustic_spaces.demos.road). The siren plays wail, yelp or two tones (--siren) at the level and directivity of a 100 W electronic siren the National Bureau of Standards measured (acoustic_spaces.demos.siren), checked byvalidation/urban_sources.py.The street’s pass-by demo drives a fire engine past (
--vehicle fire_engine): the vehicle, its siren and its air horn along one walk, each at its own sound power, with--fire-enginepicking which play (acoustic_spaces.demos.horn;police_car.jsonnow holds the car’s numbers undervehicle). The horn’s pitch, spectrum, blasts, height and directivity are assumptions, as no measured air horn of a fire engine was found.Trains and a tram in the pass-by demos, by Schall 03 (
acoustic_spaces.demos.rail): the newrailwayexample’spassby_demo.py --trainruns a regional train, an S-Bahn, an intercity or a freight train past, each as long as it is, and the street’s runs a tram (--vehicle tram), withrender_walkanddoppler.rendertakingfollowers, more sounds along one source’s walk.validation/rail_sources.pychecks a pass against Schall 03’s own sum: the same SEL within 0.01 dB in a free field, 2.9 to 4.0 dB(A) less over the example’s ground, whose reflection the engine reckons otherwise.The pass-by demos’ video, run in another format than its passes were mixed in, says which format each is in and how to match them.
Road works in the street, played by its new
works_demo.py: a breaker on an excavator, a diesel generator, a dumper and its reversing alarm, each at the sound power of Defra’s database for construction sites (the alarm SAE J994’s), checked by the newvalidation/plant_sources.py(acoustic_spaces.demos.plant).The street’s car can sound its horn (
passby_demo.py --car car_and_horn, orhornalone): a pair of disc horns as loud as Hella’s S77 (demos.horn.CAR_HORN), held against the rule for vehicles’ horns invalidation/urban_sources.py.The railway’s train can sound its warning horn (
passby_demo.py --horn blasts): the TSI’s 660 and 370 Hz notes in turn, each 109 dB(C) 25 m ahead as EN 15153-2 allows at most (demos.horn.TRAIN_HORN), or together as a chord (--horn chord), held against the rule invalidation/urban_sources.py;parameters.c_weightedgives C-weighted levels.The railway’s pass-by demo, the street’s tram and its road works can be heard indoors (
--indoors) by listeners in front of a building’s wall, as the railway’s newhouse_18mis: through a closed double-glazed window by EN ISO 12354-3 (the newacoustic_spaces.facade), the rumble of the ground’s vibration by the FTA’s general method, or both. Every listener’s caption says how hard the ground or the floor shakes and whether that is felt, and a bass shaker’s track plays it.The railway, tree belt and lakeside concert examples trace image sources to the second order, so the barrier’s and the houses’ reflections keep their partners off the ground. The railway’s trains now peak 2.2 to 2.9 dB(A) lower on the meadow, whose listener hears the barrier’s reflection, 0.7 to 1.0 lower 7.5 m from the track, and as before behind the barrier. Across the lake the evening’s weather now makes the song 7.1 dB(A) louder at the bedroom window than still air, as the house’s wall gives back what comes in off the water too (the docs’ 2.3 dB was out of date: 4.4 at the first order). Behind the tree belt’s trees, their trunks’ scattered sound now fills the ground’s dip at 500 Hz and 1 kHz: the garden there is 4.5 and 1.8 dB louder than without the wood.
The road vehicles’ sound power is CNOSSOS-EU’s Table F-1 as amended in 2021 (Commission Delegated Directive (EU) 2021/1226): the car’s at 50 km/h is 98.4 dB(A) instead of 96.0, and its passes in
roadsideandstreetare louder by about as much.Upgrading:
acoustic_spaces.demos.caris nowacoustic_spaces.demos.road(light_vehicle_power(speed, temperature)isvehicle_power("1", speed, temperature),car_soundisvehicle_sound, the coefficients are inCATEGORIES, andVehiclePowertakes the category first). The pass-by demos’ stepcarissound, and writesdemo/NAME/NAME.wav.A scene can name the
recordingits sound system plays in a demo (a file, or one of five free recordings by name, downloaded the first time), and each source thechannelit plays (left,rightormono), so that a stage’s loudspeakers play one stereo mix. In the editor: Recording on the scene’s page, Channel on a source’s.The examples with a stage, a sound system or a talker play in their demo out of the box, a song or a man talking. Their talkers now speak at the normal vocal effort of IEC 60268-16, 60 dB(A) 1 m in front, instead of the default 94 dB: the shoebox’s seat hears 60.1 dB(A) instead of 79.0, and the lecture hall’s median STI is 0.54 instead of 0.58 (one seed).
One stadium:
stadium_pais merged intostadium_concert, whose sound system plays alone for the crowd with its Monitors and Stage groups left out; its front fills keep the concert’s 116 dB, wherestadium_pa’s had drifted to 122 dB.Examples go by their names only:
examples.FORMER_NAMESis removed, andgallery_walkno longer finds the gallery.Everything a demo plays is in its example’s scene, left out of the example’s own simulations until ticked, so the editor shows it and the examples’ results stay the same to the last bit; the church demo’s places take the church’s names for
--tourand--listeners.The stadium concert and the concert stage hold the singer’s microphone (
singer_mic, a supercardioid 5 cm in front of the voice), whichfeedback.pytakes from the scene as it is; its gain before feedback moves by up to 0.8 dB in a band.The orchestra demo downloads its 61 tracks itself, by HTTP range requests out of the TU Berlin’s zips without keeping them (2.0 GB for the three movements);
--zipsstill reads zips downloaded before.fetch.open_zipandfetch.RangeFiletake the size of the blocks they ask for (block, 4 MiB as before) and keep 64 MiB of them; the orchestra’s zips are read in blocks of 32 MiB and the church’s archives in 1 MiB, which brings the church’s download from 183 MB to 127 MB.The download sizes the demos give are the measured ones, such as 313 MB for the stadium’s “Spicy Funk Cake” (450 MB unpacked, which the docs gave as the download).
Every demo with several listeners plays the ones picked for a run (
--listeners a,b), mixing and touring those alone; the demo of any scene’s--tourbecomes--listeners(--tourstill works), and the editor’s Run demo window lists the listeners to tick (demos.listeners_option,Demo.listeners).
Editor
Editor: the Scene YAML panel (View > Scene YAML) shows the scene as YAML that follows every edit, and Apply makes edited text the scene, checked and undone as any edit is (
Scene.file_text(folder, readable=True)).Paths in the editor’s Messages panel open with a click in the default app (never a script or a program), and their right-click menu copies them.
The editor’s Run demo window is simpler: the options that change what a demo plays, labelled in words (
Option.label), and the rest under Advanced (Option.advanced).The editor runs an example’s scripts that are no demos from the Run demo window too, such as the stadium’s
align.pyandfeedback.py(demos.scripts_of,demos.script_demo).The Run demo window says above Run what the run downloads and how much, until it is there (
Step.downloads).Editor: Run demo… runs the demos of an installed copy’s examples in
~/acoustic-spaces/<example>, never among the package’s files (examples.is_installed_example).Editor: sources and listeners left out of simulations (
enabled: false) start hidden in the 3D view and the plan, and their eye follows their tick.Editor: the markers of sources and listeners are all one size, in the 3D view, the plan and
acoustic-spaces view, and shrink alike where two would overlap.
Validation and docs
The validation report’s measured rooms hold four of BRAS’s, all run again with this version: the coupled rooms (CR1) and the auditorium (CR4) join the seminar room and the hall, with G compared from seat to seat and IACC against BRAS’s binaural responses (
validation/coupled_rooms.py,validation/bras_binaural.py,validation/bras_iacc_parts.py; see Validation).The validation report was checked claim by claim against its scripts and tests and corrected where it went further than they do; numbers from one-off checks are now printed by the scripts or tests, labelled as one-off checks, or left out, and every run time was measured again.
Docs: what still makes random scattering noisy where the walls are smooth (Scattering): in BRAS’s hall a few single rays hold up to 22,800 times a typical ray’s energy, and the decay times’ standard errors reach 1 to 2.8 JNDs at two of its ten pairs. Vector-based scattering stays the default.
Docs: Where the results go: the files of
simulate, the editor’s runs and exports,auralize, walkthroughs and demos, for scenes with many listeners such as the stadium.Docs: Getting started’s sample of what
simulateprints has every parameter and the standard errors, and says that the numbers move from run to run unless the scene setssimulation.seed.
4.0.1¶
Behind the archway’s wall, where the reflection off the arch’s jamb appears, the level
now steps by at most 0.08 dB from 250 Hz to 1 kHz instead of 1.37 dB, and the BRAS CR2
and CR3 rooms run at their defaults. Outdoors the early level is unchanged. Indoors it
changes at 63 and 125 Hz in every example but room_modes, whose wave solver takes
those bands, as the reflection phases now come from every band (by up to 3.8 dB, in
orchestra), and from 250 Hz up in the five examples with edges near their paths (by up
to 1.04 dB, in archway).
4.0.1 also fixes findings of the code review of 25 October 2026 (those still open are in
Known issues), among them paths through three
partitions or more, which neither the early nor the late part counted, and one the
review did not find: a closed solid standing on a surface counted as air. Of the
examples’ results, these change the apartment’s, the bedroom’s G by less than 10⁻⁶ dB
and the synthesized tails of its rendered responses, which now take the density of
reflections of each listener’s room; room_modes’s, whose wave solver keeps its highest
modes inside its stability limit, by less than 10⁻⁴ dB in G, C50 and C80 (its EDT at
250 Hz by 5 ms); every response’s last milliseconds, where the late part now takes the
whole chords of the rays through the listeners’ spheres; the echo criterion for music,
by up to 0.063 (in open_air_stage); and, as the late part of rendered responses no
longer moves ahead of 50 ms after the direct sound, the early IACC of binaural responses
by up to 0.024 (in apartment) and the echo criteria by up to 0.019 for music and 0.016
for speech (in church).
Indoors, in the bands where the early paths add in energy (from 250 Hz with the default
coherent_max_band_hz), the sound bent round edges is kept as waves (engine.clusters): each term of the edge diffraction and of the edge fade continues the geometric wave of its boundary as far as it is coherent with it over the band (on the dark side of the boundary, where that wave is gone, by the UTD’s share) and is part of its event’s diffracted wave by the rest, and a wave’s energy is that of its pieces’ pressures added up and averaged over the band, never below nothing. This replaces the cross terms between pairs of terms, which could take more than the terms carried, after which the sum stopped at zero. Behind thearchway’s wall, where the reflection off the arch’s jamb appears, the level no longer steps by 1.37 dB at 500 Hz and 1.06 dB at 250 Hz: walking 1 cm at a time it steps by at most 0.11 dB from 250 Hz to 1 kHz (0.08 dB with the edge’s image, below). Of the examples, five change: the early level from 250 Hz up by up to 0.98 dB inarchway, 0.28 dB inorchestra, 0.24 dB inconcert_hall, 0.08 dB inleaky_classroomand 0.06 dB ingallery, and at 63 and 125 Hz by up to 0.07 dB, where paths near the floor under which paths are left out come in or go (in the archway a path 40 dB down); outdoor scenes are unchanged. Over two edges with the waves in energy the level is closer to the exact solution at low frequencies (two screens 1 m apart: 1.86 dB low at 63 Hz instead of 2.74 dB) and within 0.44 dB of it from 500 Hz up (4.0.0: 0.75 dB), and the early stage of the alcove hall invalidation/edge_diffraction.pytakes about a third less time (650 ms per source and listener withfirst_orderinstead of about 1.0 s).A path that an edge cuts off on one leg, where the path over the edge would pass below the edge’s foot, takes the term of the edge’s image when the edge stands on the plane the path reflects off next to that leg: the path over the image is the path over the edge on the leg on the other side of the reflection, and its term is the cut-off leg’s. The term ended at the foot of the edge instead: behind the
archway’s wall the level stepped by 0.63 dB at 125 Hz and 0.11 dB from 250 Hz to 1 kHz between y = 6.32 and 6.33 m, where a third-order reflection that the jamb cuts off reaches the jamb’s corner just above the floor; it now changes there by 0.04 dB or less, as over the centimetres either side. Only the archway changes, by up to 0.29 dB at 63 Hz, 0.62 dB at 125 Hz and 0.13 dB from 250 Hz up; the exact solutions of the fade invalidation/edge_fade.pyare matched as before.In the bands where the early paths add in pressure, a face’s reflection takes its phase from an admittance fitted to the face’s absorption in every band, as the ground outdoors already did, instead of in those bands alone, so the level in a band no longer depends on how many bands add in pressure (
coherent_max_band_hz). Behind thearchway’s wall the level at 125 Hz is -27.29 dB with the default and -27.28 dB with every band in pressure, where 4.0.0 gave -27.17 and -27.91 dB; walking along the wall the two stay within 0.08 dB at 63 and 125 Hz, where paths near the floor under which paths are left out come in or go. Against porous layers of known impedance (tests/test_coherent.py), the phases at 63 and 125 Hz from a fit to every band are less than half as far off as those from a fit to the two bands alone. The early level at 63 and 125 Hz changes in the rooms: by up to 2.8 and 3.8 dB inorchestra, 2 to 2.7 dB at 125 Hz inapartment,archway,concert_hall,listening_roomandstudio, and less in the others; the parameters of BRAS CR2 change by 0.1 JND or less. Outdoor scenes, the bands the wave solver takes (room_modes) and the bands from 250 Hz up are unchanged. The fit takes up to about 0.5 s per material, once per process.Re-weighting a run for other absorption (the editor’s results after a material’s absorption changes,
Response.reweighted) turns the phase of each early reflection off a changed material where the paths add in pressure (up to 125 Hz indoors by default, every band outdoors), as tracing again does. It kept the phases as traced, which a reflection factor fitted to the absorption sets: with thestudio’s side panels 0.2 more absorbent, G, C50 and C80 at 63 Hz came out up to 0.98 dB from those with the new phases. The paths keep the angle of each reflection (EarlyReflections.cosines), andResponse.reweightedtakes the absorption as traced and as re-weighted for (absorption=(old, new)), which it needs where a path reflects off a group whose absorption changes. As for their energy, the faces of an edge that a path bends round keep what they did to its phase as traced.calibrate’s results are unchanged: its fit within a round keeps the traced phases, and each round traces again.A room with a curved surface of many facets, whose events at the edges have hundreds of thousands of terms, no longer runs out of memory in the early stage: the BRAS CR3 room at its defaults peaks at 10.0 GiB (11.7 GiB in third-octave bands), where 4.0.0 asked for 16.4 GiB at once, and CR2 at 1.1 GiB, where 4.0.0 asked for 19.6 GiB. The edge fade keeps the pieces of its waves in single precision, leaves out those with nothing in them before it keeps them, and hands back to the system what they leave free in the C library’s heap (glibc’s
malloc_trim).A path that reflects and passes through three partitions or more (faces that reflect and let sound through) was counted by neither part: the early stage follows a path through two partitions at most, and the stochastic late stage left out the specular part of every ray up to
early.max_orderreflections, wherever it went. In four rooms in a row behind partitions that let 30 % through, the early stage’s ray tracing at order 3 missed 18 of the 62 paths through three partitions into the fourth room, 21 % of its early energy. The late stage now counts the partitions each ray passes and keeps the reflected paths through more than two, and the early stage leaves those out: the fourth room’s response is the same at order 3 as at order 0, and ray tracing finds the same early paths as image sources in every room. With the statistical late method, which takes the early paths’ energy out of its own, the early stage keeps them. In theapartment, the only example with partitions, the bedroom keeps 70 of its 76 early paths and its G changes by less than 10⁻⁶ dB at the same seed. From order 10 up, one ray of the early stage’s ray tracing also kept only 64 of the branches it forks into at partitions (66 at order 10); it keeps them all.A closed solid that touches the rest of the model at no edge (a pillar, a platform, a cabinet) is placed in its room from points just off the largest faces of up to six of its planes, which vote by those faces’ areas, instead of off its largest face alone. Where that face lay on another surface, as a platform’s underside on the floor, the point was outside every room and the solid’s inside counted as a room of its own: a 1 m cube on the floor of a 3 m room gave 28 m³ instead of 26, a platform 1.5 x 1.5 x 0.5 m 28.1 instead of 25.9. Every example keeps its volume, and finding the cells of a city of 10,000 buildings takes as long as before.
A source or listener inside a closed solid (a column, a block, a thick wall) is reported as such, and
simulaterefuses to run, where the solid’s inside became a room of its own: a listener inside a 1 m³ block on the floor of a 3 m room made the volume 27 m³ instead of 26, and the run only warned that the listener was in another room than the source. Such a space still holds air when a surface that lets sound through joins it to a room (a booth with glass walls).inspectmarks the positionIN A SOLID. When no source or listener stands in the model’s rooms (walls drawn as a closed shell inside another), the spaces they stand in are the rooms, as before.A source or listener on a surface (within 1 µm of it) is reported as such, and
simulaterefuses to run: the inside checks can say either there, and a listener on the shoebox’s far wall counted as inside, with a detector sphere of 1 mm. An opening, such as an outdoor scene’s sky, has air on both sides and does not count. A listener closer than 5.6 cm to a surface, an opening included, gets a warning where the late part is traced with detector spheres sized for the room: kept clear of the surface, its sphere is under a tenth of the smallest such one and catches a hundredth of the rays or fewer. A centimetre from the shoebox’s wall, G at 1 kHz came out 1.5 dB apart between two seeds and T30 0.97 and 1.40 s, against 1.24 and 1.27 s a metre from it.validate --strictexits with status 1 on warnings too (a model that is not closed, for one), for scripts that should stop there.A material with
transmission: 0lets nothing through when the rooms are found either: in walls drawn as a shell, a room joined to the source’s only by such a wall held air too (rooms of 48 and 72 m³ gave 120 m³ instead of 48).What a surface lets out of the rooms counts as absorbed for the rain of the late part (the energy each reflection sends straight to the listeners) as it did for the rays: on the CPU the rain went out of a room through one surface and back in through another, which no ray can do, so two houses across a lane whose facing walls let sound through heard each other’s rain. The GPU’s rain stops at every surface, as before. The examples have no such path: their results are the same to the bit.
The late part’s last milliseconds take the whole chord of a ray through a listener’s detector sphere, or through a coverage map’s cell, where the ray’s time runs out as it crosses it: a ray’s legs now go on past its time by twice the largest sphere’s radius (or map cell’s half diagonal) without reflecting there, and a chord counts if its closest approach (a map cell’s: its middle) is in time. The chords were cut short where the time ran out, and one whose closest approach came later was counted at the end: in the example box the last 1 to 3 ms came out up to 38 % low with 1 m spheres and no rain, and the last millisecond 17 to 23 % low with the default diffuse rain and 0.5 m spheres, on the CPU and the GPU alike. The total energy moves by 2.1e-4 at most. The tracer’s reference results (
tests/reference/tracer.npz) are written again: only their last slots change.A rendered response’s late part, spread over 10 ms from 50 ms after the direct sound so that a single strong ray does not become a click (
synthesis.smooth_energy), now spreads each slot only over the part of its window from those 50 ms to the end. The window centred on each slot moved energy up to 5 ms ahead of that time, into C50’s early part, and lost some past the end: at the examples’ own settings it moved 0.03 to 5.2 % of the late energy after those 50 ms ahead of them (5.2 % instudio), which raised C50 of the energy the responses are made from by up to 0.45 dB (gallery, round the corner). The parameters a run reports come from its echograms, which are not smoothed, and do not change. Measured on the rendered responses, the early IACC of the binaural ones moves by up to 0.024 (apartment, television to bed), the echo criteria by up to 0.019 for music and 0.016 for speech (church, altar to the back of the nave), and the late IACC by 1.4 × 10⁻⁴ at most.Where partitions or openings divide the model into rooms, the statistical method’s mixing time (where its early paths end) and the density of reflections in the synthesized tail come from the volume of the listener’s room (a map’s mixing time from each point’s), not from all the rooms together.
Room.volume_atgives it, and a response’svolume_m3is now its listener’s room’s. The stochastic tracer’s detector spheres stay sized for all the rooms, which the rays spread over: sized for a side room’s own volume, they gave the same results with half as much again spread between seeds.The model’s repair drops vertices that no face uses. A stray one moved the point that places a separate shell, set off its face by a millionth of the vertices’ extent: one 1000 km off set it 1 m off, and a slab’s inside became a room (28.6 m³ instead of 25.4). Every format now refuses coordinates that are not finite and faces that name a missing vertex, where only the OBJ reader refused the first.
The model’s repair drops a face given twice (the same corners, turning the same way), counted with the degenerate faces it drops. The copy left each edge of the face shared by three faces, and the room open with no volume: one wall triangle given twice in a box left 13 faces with no room on either side. Two faces with the same corners turning opposite ways both stay, as where two closed models touch.
packwrites the zip as<name>.zip.partand puts it in place only once it is complete, so a pack that fails leaves an earlier one of the same name as it was; before, the earlier one was emptied when packing began and deleted when it failed.Unpacking a pack first checks that the folder has room for its files, at the sizes the pack records (as much as reading them gives), and writes each file as
<file>.part, which takes its name once its hash matches: a damaged file is no longer left in the folder. A file that is encrypted or compressed in a way Python’szipfilecannot read is refused as damaged, where it ended in a traceback.A key written twice in one mapping of a scene file or an array’s YAML file is refused (
found the key 'materials' twice): PyYAML keeps the last, so a secondmaterials:dropped the first without a word. A key that a merge (<<:) brings in may still be given again.Result files of names that give the same file name, such as
A/BandA_B(anything but ASCII letters, digits, dots, dashes and underscores becomes_), or that differ only in case, get-2,-3and so on after the later ones, where the second response’s WAV file overwrote the first’s. Two binaural outputs through HRTF files of the same name, andwalkthrough’s files per listener, are kept apart the same way.The results of
simulateand the editor’s Export are written into a hidden folder in the results folder and moved into place once all of them are,parameters.jsonlast: a run that fails while writing (a full disk) or an export that is cancelled leaves the folder as it was. Before, it left the new run’s first files beside the earlier run’s report and other files.simulatechecks its--outfolder before the run: an existing file there stopped it with a traceback only once the run was done, and now stops it at once, as a folder it cannot make or write in does (the folders it made are removed again when the run fails). The responses are rendered and written one at a time, on the run’s threads, instead of all held until the end: writing 48 responses in mono, binaural and third-order Ambisonics on 4 threads peaked at about 1,280 MiB, and now at 725 to 821 MiB whatever the number of responses (8, 24 or 48; it varies with how the threads’ renders overlap).write_wavwrites a few MiB at a time. The files are the same to the byte but with--normalize, whose files are scaled once all are written, by one rounding to single precision more (one unit in the last place at most).parameters.free_field_reference, which the--free-fieldoptions read, takes the bands with 0.2 s of silence on both sides of the response. It lost what the zero-phase filters spread before the first sample: of a response that starts with its direct sound it kept 0.50 of the energy at 63 Hz to 0.71 at 8 kHz, so G against it read up to 3 dB too high.octave_bandsandsplit_bandsreturn signals as long as their input, which cannot hold that spread; their docstrings now say how much is lost. The same loss at the start of a rendered impulse response is a known issue.parameters.from_impulse_responserefuses areference_energythat is neither one energy nor one per octave band: 24 third-octave references were cut to their first 8 and taken for octaves, and a list of one raised anIndexError.The echo criterion for music filters the response with one band-pass over the 1 and 2 kHz octaves (two octaves about 1.4 kHz, as Dietsch and Kraak have it) instead of the two octave filters added, which rise 3 dB where they meet: an echo near 1.4 kHz reached the 50 % mark 2.0 to 2.2 dB quieter than one at 1 or 2 kHz, and now all three reach it at the same level within 0.25 dB. Over 24 responses in seven examples the criterion moves by 0.063 at most (
open_air_stage, 300 m downwind: 0.225 to 0.288). The criterion for speech, and both with--filters crossover(whose bands add up flat), are unchanged to the bit.Absorption or scattering outside 0 to 1 read from a preset, a CSV file or a CATT or ODEON library is still taken as 0 or 1, but the run now warns, naming the material, where the values came from and each band out of range (
Material.clipped). Values written in the scene were already refused.No path arrives before it leaves the source: a moment of turbulence keeps every path’s delay at zero or more, and the echogram, calibration and the statistical late method count a path of negative delay in the first time slot instead of the last. No run reached it (in strong turbulence a path moved by at most 0.5 % of its delay).
An error in one of the native code’s worker threads (running out of memory, say) is raised in Python instead of ending the process.
The stochastic tracer counts a ray’s reflections off each group of faces (for re-weighting) in 32 bits instead of 16, which went back to 0 after 65,535 reflections off one group: in a 1 m cube of one material, a ray reflects that often in about two minutes.
The docs said results do not depend on
threads. They do, by rounding: each thread sums its share of the rays, and the shares add up in another order (in the shoebox example, 7 × 10⁻¹⁶ of the late energy’s peak between 1, 2 and 4 threads; the early paths are the same to the bit).Convolver.processtakesout, an array to write the block into, andplaykeeps one such buffer, so the sound card’s callback no longer makes new arrays the size of a block;HeadTrackedBinaural.processtakesouttoo and keeps its buffers.The wave solver’s update takes its Laplacian times 1 − 2⁻¹⁸ (sound 2⁻¹⁹ slower), which keeps the grid’s highest modes inside the stability limit in single precision. At the limit itself they sat on its edge, where a kick grows linearly: the 7-point stencil’s checkerboard 2⁻²³ inside, which rounding can take over it, and the 19-point stencil’s modes alternating along two axes exactly on it. Kicked, they grew to 7,000 and 20,000 times the kick in 20,000 steps; 2⁻¹⁶ inside, they swing within 259 times it over 300,000. No room was seen to grow: its walls and the grid’s edge keep its highest mode further inside (a rigid box ran 200,000 steps). The 7-point stencil’s weights, rounded down, also made a room’s uniform pressure swing slowly (0.3 Hz with 10 cm cells), which reached into the bands: in a duct, the reflection off an end whose admittance resonates now follows the grid’s theory within 0.001 instead of 0.003 from 30 to 250 Hz. The wave solver’s validation tables move by one in their last digit at most.
The wave solver checks its grid against its limit of 40 million cells before it makes the weather’s fields for it, so a grid too large outdoors in wind or a temperature gradient is refused with the message that says what to lower instead of running out of memory; and round an open model it lets its first grid go before it makes the one with the matched layer (for a 30 x 20 m open floor at 250 Hz, 151 MiB at the peak instead of 209).
engine.wave.wave_responseschecks what it is given, as the scene’s schema does: a mistypedschemeran the isotropic stencil and a mistypedwindthe effective one, a length of 0 ran with no samples, and fewer names than listeners left listeners out. A model whose every face is an opening is refused as having no surface that reflects, where the solver stopped withBvh: the mesh has no faces.The wave solver on a GPU checks the matched layer’s state against what the GPU binds in one buffer, as it did the field: round an open model it is the largest buffer (4 to 7 times the field’s round the tests’ 5 x 4 m ground, with cells of 6 to 30 cm), and a grid whose layer the GPU could not bind stopped with wgpu’s validation error once it was built, instead of being refused with what to lower. The pressures at the listeners go to the GPU’s buffer as many steps at a time as it binds.
The wave solver’s recording of the pressures at the listeners is no longer copied once more at the end of the run: with 60 listeners and 4 s responses at 63 Hz the run peaks 12 to 13 MiB lower, the recording’s size. The results are the same.
Below the default 8 cells per wavelength (
simulation.wave.points_per_wavelength, which takes 5 to 20), the scene’s warnings say how much slower the wave solver’s shortest waves travel, at the top of its highest band along the grid’s axes, where both stencils are slowest: at 5, 1.85 % with the isotropic stencil and 4.56 % with the standard one, against 0.67 % and 1.74 % at 8, so the modes there come out that much low. Each run’s details give it for the grid the run took, aswave_slow_percent(in wind and temperature, for the slowest sound on it).engine.wave.slow_byandgrid_phasegive the stencils’ dispersion relation, which a tube’s modes at 12, 8 and 5 cells per wavelength follow within 2e-5.A run’s memory is estimated before it starts (
engine.run_size): the late part’s energy at every listener or map point for every source, and while a source is traced, the tracer’s copies on the CPU’s threads or its buffers on a GPU, the worst of the deviceslate.devicemay take on this computer. A simulation or a coverage map that would need more than the computer’s memory (or its container’s limit, on the process’s own control group or one above it) is refused with a message that says what to lower, andRoom.warnings()(the editor’s messages) notes one that would need more than half. A map is checked again once its first source point is traced, with the early paths found there at every point.performance.threadsabove the computer’s cores run as its cores. On the CPU the estimate came within 1 MiB of the late part’s measured peak at 1, 4 and 64 threads; see Memory for what it leaves out.The late part takes less memory, with the same results: the stochastic tracer keeps its threads’ copies of the results within 256 MiB together (one thread at least), where only those of a map and of the moments had a limit, each its own; a run traced on the CPU and the GPU together adds each part to the sums as it comes in instead of keeping every part; the two batches of the weight window are mixed in place and the bands joined as each is traced; the GPU’s sums are read back one array at a time; a line or surface source’s points are summed as each is traced; and a source with a feed of its own lets its late part go once its responses have their copies.
stadium_paat 8 s peaks at 547 MiB instead of about 1,030 MiB. The results are the same bit for bit on the CPU, in maps, and on the GPU without the weight window (with it, and on both, they vary from run to run within the noise of the rays, as before).CI checks types with mypy, on every module but the 70 of 116 that did not pass when the check began (
pyproject.tomllists them; each leaves the list once it passes), runs the tests on a build of the native module with AddressSanitizer and UndefinedBehaviorSanitizer (the CMake optionACOUSTIC_SPACES_SANITIZE), and measures the tests’ coverage of the Python code. Thedevextra and group install mypy and PyYAML’s types.The validation report’s measured rooms (BRAS CR2 and CR3), made before 3.0.0, and its cross-check with pyroomacoustics are run again with this version, with their run times and memory measured on 4 cores (see Validation). The seminar room’s mean differences from the measurement move by 0.3 JNDs at most. The hall’s EDT and Ts come closer (with the fitted absorption EDT 2.1 JNDs off on average instead of 2.6, Ts 0.9 instead of 1.3), but its EDT at 63 and 125 Hz, 23 and 19 % long before, comes out 19 and 13 % short: with edge diffraction its early part there is 6.6 and 5.8 dB stronger than with that code, 3.1 and 4.4 dB of it from the fade at edges that 4.0.0 added. The report said the diffuse box’s C80 from the energies came within 0.7 dB of Barron’s revised theory; it is 0.6 to 1.0 dB below, as it was with the code of the report.
The edge fade is walked in CI (
tests/test_walks.py): thearchway’s whole line behind its wall, thegalleryround its corner and thestreet’s garden round its four largest steps, each held to the steps it makes now.validation/walks.pymakes the whole walks, with the fade and without it. The validation report’s table of the steps along the street is made again with it, 1 cm apart round every step among the four largest for its listener in a band or in dB(A): 20 places instead of 11, among them the largest in dB(A), 0.52 dB(A) (1.13 dB at 1 kHz) in the garden with the car at x = -41.39 m, where a term reaches the end of the barrier’s top.The validation report checks the edge diffraction and the fade against the exact solution in three more places (
validation/edge_diffraction.py,validation/edge_fade.py): past the corner of a thin screen withdirect, behind a barrier 3 m high on a hard ground withfirst_order(one_sideandboth_sides) and behind a barrier 1 m thick withsecond_order.btm_firstthere takes the ends of a finite edge.Known issues, measured: past the corners of edges, which the UTD has no terms for, the level steps (by 2.8 dB at 63 Hz with
directpast a thin screen’s corner, where BTM changes by 0.15 dB or less; along thestreetby up to 1.42 dB in a band and 0.52 dB(A)). Behind a barrier on a hard ground withfirst_order, the path over the top with the reflection off the ground behind it is left out where the reflection before the barrier is found: the level steps by 0.35 dB at 63 Hz there and is 0.33 dB low above (both_sidesis within 0.07 dB of the exact level from 125 Hz up). A fix changes BRAS CR2’s early level by up to 0.8 dB at 63 Hz, so it waits for a run of BRAS with it. Withsecond_ordera path over two edges takes no reflections, so behind a thick barrier on a hard ground the level is 1.7 to 5.7 dB low from 500 Hz up.Known issue, documented: with vector-based scattering (the default) the late part carries the specular share of the first orders with more energy than the exact paths (a quarter more for orders 1 to 3 near the middle of a small box), so the total depends a little on the early order: within 0.04 dB of exact paths up to order 9 at the default
early.max_order: 3, 0.15 to 0.2 dB higher withmax_order: 0. Random scattering keeps within 0.03 dB.Known issue, measured: outdoors, where the echogram’s pairs of paths from 1 kHz up would take a time slot below zero and it is set to zero, the energy added comes to at most 0.009 dB of a band’s response over the nine outdoor examples’ 419 pairs of a source and a listener, and 0.04 dB in the lateral figures of eight.
4.0.0¶
Outdoor acoustics: scenes on a ground under the sky, with buildings and barriers, terrain and buildings from a map, grounds of finite and of extended impedance, wind and temperature gradients, turbulence, vegetation, weather zones and classes with levels over a long term, moving sources with Doppler, whole cities and a wide-angle parabolic equation, on the GPU too, validated against a fast field program. Reflections that fade in and out at the edges that cut them off, coherent arrays element by element, groups of sources and listeners, demos of the examples (a car driving past with Doppler among them) and of any scene with files of your own, walks along keyframes with every mover at once on one clock and the editor’s Movement panel, HRTF sets to download, the checks of ISO 3382-2 on every decay time, the wind drawn in the editor, and packs that carry a scene with its files to another computer, with a container image that runs them headless on a GPU.
Indoor scenes give the results of 3.2.3 but where edges diffract (Reflections that
fade at edges) and for what Fixes lists: every binaural output changes (the
treble an HRTF between the measured directions lost is back), a response that ends in
silence no longer has decay times read from the drop after its last sound, auralize
and walkthrough mix sources at their sound power (--equal-power keeps the old mix),
and the last bits round alike on every platform.
Outdoor scenes
geometry.groundmakes a scene outdoors: a flat ground under an open sky, which lets all sound out, withbuildings(footprints raised to a height, open underneath),barriers(thin screens along a path) and a model file’s objects standing on it. The ground reachesmarginpast them or over a setextent, and the sky liesskyover it ormarginover the highest of them. Surfaces are named after each building (<name>_roof,<name>_wall_1…) and by kind (walls,roofs,buildings,barriers,ground,sky).Nothing reflects off the sky, no edge of the sky or on the ground diffracts and the ground’s size scatters nothing, so the ground reads as endless; the inside corners of buildings do not diffract either.
Sources and listeners inside a building, below the ground, above the sky or past the ground’s edge are reported as problems.
Outdoors,
ir_length_scounts from the latest direct sound, so far listeners are not cut off; the statistical late method, which needs a diffuse field, stops with an error; coverage maps cover the ground and leave out points inside buildings.The editor’s Geometry form has an Outdoors group (ground height, margin, sky, buildings’ and barriers’ heights); the 3D view, the plots and the Blender export leave the sky out;
inspectand the status bar describe the ground instead of a reverberation time.New example:
street, traffic in a street between two buildings, heard on the pavement, on a balcony and in a garden behind a noise barrier.Rooms are unchanged: indoor scenes give the same results, bit for bit.
Ground impedance
A material’s
impedancemakes it a porous ground: a Nord2000 ground class (snow,forest_floor,loose_soil,grass,compacted_soil,gravel,asphalt,water) or a flow resistivity, by Miki’s model (the default), Delany and Bazley’s or the two-parameter model, optionally as a layer on a hard backing. Its absorption is its random-incidence absorption, in third-octave bands.Reflections off it take the spherical wave’s reflection factor (Weyl-van der Pol), with the phase and the ground wave near grazing incidence that make the ground dip: within 0.05 dB of the exact solution over grass and harder grounds, 0.2 dB over snow and 0.4 dB over a layer of snow 10 cm thick, outside dips more than 20 dB deep.
Ground zones (
geometry.ground.zones): parts of the ground with materials of their own, such as a road across a meadow. A reflection near where two grounds meet takes each one’s factor, weighted by the share of its Fresnel zone each covers, band by band.Outdoors the early paths add in pressure in every band by default (
coherent_max_band_hz: 8000; a room takes 500 at most), and the paths of each point of a line or surface source add in pressure, the points in energy.The wave solver fits a ground to its impedance itself, phase and all, with branches without mass: within 0.12 dB of the exact solution over grass from 40 to 315 Hz.
The editor’s material page has a Ground impedance group; the Outdoors group lists the ground zones; In pressure up to goes up to 8000 Hz outdoors.
inspectshows each material’s impedance and the zones;--coherent-bandtakes 1000 to 8000 outdoors.calibrateleaves materials with an impedance alone.New example:
roadside, traffic on a road between a meadow and a paved yard, with the ground dip over the grass. New validation:validation/ground_reflection.py.
Wind and temperature
atmosphere(outdoor scenes): the wind and the temperature over the ground, which bend sound down (downwind, a clear night) or up into shadow zones (upwind, a sunny day). Presetsneutral,windy,sunny_dayandclear_night; the wind’s speed, direction, height and log, linear or constant profile; a temperature gradient and a log term near the ground; the ground’s roughness; or a measured table of heights, temperatures and wind speeds.The direct sound and the reflections off the open ground are the eigenrays of the effective speed of sound, curved paths with exact travel times, levels from the ray tubes’ spreading (capped at caustics) and the phase lost at each caustic: several downwind, none in a shadow. Paths off façades bend in their unfolded plane; paths over edges fall into the shadow as a creeping wave.
Below 500 Hz, and up to 2 kHz in and near a shadow, the sound over the open ground comes from a Green’s function parabolic equation over the grounds’ impedances, with Greene’s starting field: within 0.25 dB of the exact solution in still air over grass and harder grounds. It runs on
simulation.performance.threads.The late rays follow their arcs through the layers; the animation’s particles,
view --pathsand the Blender export curve too. The direct sound’s time in the results is the bent path’s.The editor’s scene page has a Wind and temperature group with Show the profile;
inspectdescribes the atmosphere and the speed of sound at a few heights;acoustic_spaces.viz.plot_atmospheredraws it.New example:
open_air_stage, a stage heard 100 and 300 m away across a park, downwind and upwind in a 5 m/s wind. New validation:validation/outdoor_propagation.py, a wind turbine heard 178 to 845 m away (the Hannover data, downloaded on first use): 1.7 to 2.7 dB per band on average in light winds, as close as the authors’ own model.Scenes without an atmosphere, or with one in which sound goes straight, give the same results as before, bit for bit.
Turbulence
atmosphere.turbulence: the eddies of the wind and of the temperature, by their structure parameters C_T² and C_v² and the size of the largest eddies, with a von Kármán spectrum; presetsweak,moderateandstrong. Also in still air.Each two early paths of a sound add in pressure with their coherence factor, from the Markov approximation along both, reduced for paths that differ only by reflections off the ground as a parabolic equation marched through random eddies has it: the ground dip fills in at long distances and high frequencies, within about 1 dB of that parabolic equation on average over eight grounds and heights. The echogram, the parameters and the coverage maps take this average; in the parabolic equation’s bands the eddies blur its sound towards the rays’ added in energy.
Eddies scatter sound into the shadows of the wind and of the temperature: once from the air over the shadow’s edge (the Born approximation), and at high frequencies by shaking the refraction near the ground: within about 1 dB of a parabolic equation through random eddies on average. 300 m upwind in the
windypreset the bands from 1 kHz up are about 19 dB below those downwind with moderate eddies, 60 without.An impulse response is one moment of the eddies (drawn with the seed), each path delayed band by band;
auralize, the editor’s Listen andwalkthroughgo through a new moment every sixteenth of the time the eddies take to drift past (at most 20 times a second), crossfaded, so the sound flickers.The editor’s Wind and temperature group has Eddies (none, a preset or custom) with the structure parameters and the eddies’ size;
inspectdescribes them.The
open_air_stageexample has moderate eddies. New validation:validation/turbulence.py, against a parabolic equation through random eddies;validation/outdoor_propagation.py --turbulenceadds eddies to the Hannover cases.Scenes without turbulence give the same results as before, bit for bit.
Vegetation
geometry.vegetation(outdoor scenes): belts of trees and woods (a footprint raised to a height), hedges (a path of a width, mitred at its bends) and lone trees (a spheroid crown), which sound passes through. Each holds foliage, ISO 9613-2’sdensepreset (the default, all absorbed) or an attenuation per band with an albedo and a mean cosine of scattering, and stems, upright rigid cylinders by their number per m² and their diameter, which scatter by the exact series and only sound’s horizontal share. Overlapping volumes add up, and a volume may run past the ground’s edge.The early paths lose what the plants take along each leg (Beer-Lambert), the parabolic equation’s sound what its direct path loses. The late rays absorb along their way and scatter in the plants at random: foliage by Henyey and Greenstein’s phase function, with its share sent straight to each listener, stems sideways, keeping the slope, with what the next leg carries through each listener’s sphere averaged over the turns; in several bands at once each band keeps its own expectation. The rain from every reflection loses what the plants take on its way. Coverage maps take both. The wave solver’s bands leave vegetation out.
The editor’s Outdoors group lists the vegetation (its shape, what fills it, its heights, and Remove); the 3D view,
viewand the Blender export draw it in see-through green, coverage maps outline it, andinspectgives what sound loses per metre through each.New example:
tree_belt, traffic on a road heard behind a belt of trees and across an open meadow. New validation:validation/vegetation.py, the late rays against a separate particle simulation through a slab of foliage or stems: within a few tenths of a percent.Scenes without vegetation give the same results as before, bit for bit.
The wave solver in wind and temperature
Outdoors, in an atmosphere whose wind or temperature changes with height, the wave solver’s speed of sound changes from cell to cell: the still air’s at each cell’s height, plus the wind’s component along the horizontal from the source, the effective speed of sound the rays and the parabolic equation take. The grid is fine enough for the slowest sound and steps in time for the fastest; each cell weighs its neighbours by the square of its speed over the grid’s, its walls by that speed (their admittance is relative to the local rho c) and the source by its square, and so does the matched layer. The flow itself (the linearised Euler equations) stays out.
The run’s details give the slowest and the fastest speed of sound on the grid (
wave_speed_min_m_s,wave_speed_max_m_s).New validation:
validation/wave_outdoor.py, the 63 and 125 Hz bands over grass, 50 to 200 m from the source, in still air, downwind and upwind, on a sunny day and on a clear night, against the parabolic equation: within 0.7 dB at 63 Hz and 0.4 dB at 125 Hz, 0.2 dB on average.Indoors, and outdoors in still air, the wave solver gives the same results as before, bit for bit.
Outdoor results
Outdoors, every response and map point gives its level relative to free field, per octave band and A-weighted: how much louder the source is there than the same source as far away in free field, with the same air; with the sign turned, the excess attenuation.
simulateprints it,parameters.jsonholds it (re_free_field_db,re_free_field_a_db), maps have the parameterre_free_fieldand the editor’s Results panel the row Re free field.simulation.insertion_loss_ofnames buildings, barriers or vegetation (--withouton the command line, forsimulateandmap; an Insertion loss list on the editor’s Simulation page): the scene is run again without them, with the same seed, and the results carry their insertion loss at every listener (insertion_loss_db,insertion_loss_a_db) and map point (insertion_loss,insertion_loss_a), the level without them less the level with them; the report says what was left out (removed).Scene.without(names)is the scene without them, andacoustic_spaces.engine.insertionruns both.Outdoors, G of a response that holds no decay to read (an open field, a lone barrier) is that of the energy it holds instead of undefined, and so are the levels and the STI that take it; the other parameters stay undefined. Indoors nothing changes.
New validation:
validation/insertion_loss.py, a barrier’s insertion loss against the exact solution for a half-plane (within 0.07 dB over an absorbing ground) and the level over grass against the exact solution for a ground (within 0.15 dB, up to 1.1 dB in the ground dip below 1 kHz, which the band filters fill), and a barrier over a rigid ground against the four paths over it (below).Scenes without
insertion_loss_ofgive the same results as before, bit for bit, except that G outdoors is now defined where it was not.
Reflections on both sides of an edge
simulation.early.diffraction_reflections: both_sides(--diffraction-reflectionsforsimulateandmap; Reflections over an edge on the editor’s Simulation page) lets a diffracted path reflect both before and after the edge: off the ground on the way to a barrier and again behind it, the fourth of the four paths over a barrier on the ground, which carries a quarter of the sound over it on a hard ground. Its reflections take the materials’ coefficients and phases as any other path’s, and it counts as a diffracted path of order 2 (its two reflections) where paths are drawn, exported and reported.Validation (
validation/insertion_loss.py): a barrier on a rigid ground against the four paths by the exact solution, the exact solution read through the engine’s own band filters below 1 kHz: within 0.15 dB at every listener and band withboth_sides; withone_sidethe insertion loss is 0.3 to 0.8 dB too high from 1 kHz up and up to 3 dB off below.one_sidestays the default, so existing scenes give the same results, bit for bit.
Foliage from the leaves’ size and density
foliage: {leaf_area_per_m3: F, leaf_width: a}on a vegetation volume gives its leaves instead of a preset or measured rates: their area (one side’s) per cubic metre and their width, the diameter of a disc of a leaf’s area. The engine works out what they take in each band from thin rigid discs turned every way alike (engine.vegetation.disc_scattering,leaves_medium): a dipole, from the air a small disc holds back, 16 (kr)⁴ / (81 π²) of its area; twice the mean projected area of a large one, half the forward shadow and half reflected every way alike; the two bridged smoothly in between. They absorb in the viscous and thermal boundary layers on their faces (leaf_absorption, Morse and Ingard’s rigid wall), a few per mille of their area. The scattered share and its mean cosine per band replacealbedoandasymmetry. The editor’s vegetation list andinspectshow such leaves; the late rays, the early paths and coverage maps take them as any other foliage.New validation:
validation/leaves.py, the disc’s cross-section against its closed forms at both ends, the absorption against the wall’s in closed form, and a wood of leaves 50 m deep compared with ISO 9613-2’s dense-foliage table: leaves 5 cm wide at 0.1 m² per m³ take about what the table gives at 4 kHz (3 to 3.5 dB against 4.5), more at 8 kHz and next to nothing below 1 kHz, where the table’s 1 to 3 dB are not what leaves scatter.validation/vegetation.pyruns the late rays through such leaves against its particle simulation too.The
densepreset,noneand measuredattenuationgive the same results as before, bit for bit.
Terrain and buildings from a map
ground.terrainmakes the ground uneven: its heights on a square grid, from an Esri ASCII grid or a table of x y z points (file;gdal_translate -of AAIGridturns a GeoTIFF into the first) or given in the scene (heights,origin,spacing), bilinear between the nodes;ground.map_originis subtracted from a map’s coordinates, andground.heightadded to the heights. The scene’s ground is the grid sampled over its extent (the whole grid unless an extent is set) and cut into triangles, with the sky walls rising from its edge; buildings, barriers and vegetation stand on it however it runs, and ground zones take its triangles. Files in degrees of longitude and latitude are refused.inspectdescribes the terrain and gives each source’s and listener’s height over the ground.geometry.buildings_fileadds the buildings of a GeoJSON file:PolygonandMultiPolygonfootprints raised to the height a property gives (height,default_height), named by a property (name), some left out (without); holes are built over and reported. The editor lists them after the scene’s own buildings and removes one intowithout.Over a terrain the early sound reflects once off the mean plane of the ground’s profile between the source and the listener (CNOSSOS-EU’s and Nord2000’s ground reflection, with the ground’s reflection factor under the reflection point), not off the terrain’s facets (
engine.terrain), and bends over the crests the line of sight cuts through (engine.crests): the string pulled taut over the profile, one wedge of the uniform theory of diffraction per crest with the ground’s reflections off the profile’s segments on both sides in its terms, Fresnel-zone weighted as in Nord2000 (a term of the wedge’s coefficient near the crest, a path of its own far from it), the equivalent edge of a flat or rounded top with ISO 9613-2’s thick-screen factor over the ground within an eighth of a wavelength of the string (nothing for a ridge), and in sight the nearest crest’s wave with the transition at its boundaries and the reflection off the slope before it, so that the level is continuous across the edge of a hill’s shadow and ripples in the lit zone as the exact field does. The late rays bounce off the terrain’s triangles. Edges buried in a terrain do not diffract.New validation:
validation/terrain.py, a flat terrain against the flat ground (the same paths to rounding), a plane slope against the image source in the tilted plane, a ridge against the exact solution for the wedge (within 0.2 dB in every band), a smooth hill against the thin screen’s attenuation (Maekawa’s formula over a soft ground, ISO 9613-2’s over a rigid one, where the flanks’ reflections add) and the level’s steps as a listener climbs out of its shadow, and the cost of a terrain 2 km across (see Validation).New example:
hillside, a works yard at the foot of a hill heard in the hill’s shadow, on its slope and on its crest, over a terrain from a grid of heights with two houses from a map.Scenes without a terrain or a buildings file give the same results as before, bit for bit.
Outdoor scenes in the editor
File > New outdoor scene starts a flat meadow 200 m across under an open sky, with a source and a listener 50 m apart, to build on.
The Geometry form’s Outdoors group draws new parts in the plan view: Add building, Add zone and a belt of trees as outlines closed on their first corner, Add barrier and a hedge as paths ended with Enter or a double-click, and a lone tree with one click; each takes the usual sizes (a building 10 m high, a barrier 3 m, a belt 12 m, a hedge 2 m high and 1.5 m wide, a tree’s crown 8 m across from 2.5 to 10 m up, dense foliage) to change afterwards, and Remove takes it out again. The plan shows the vegetation in dashed green and the zones in their materials’ colours.
The ground’s Extent is set in the form (or left to the margin), Terrain from a file and Remove terrain set and drop a terrain, and Buildings from a map adds a GeoJSON file’s buildings after asking which of its properties hold their heights (or one height for all) and their names, with Remove the map’s buildings to drop the file;
acoustic_spaces.geometry.terrain.footprint_propertiesreads what a file offers.A saved scene refers to its terrain and map by paths relative to the scene file, like its model file.
Validation against a fast field program
validation/ffp.py: a fast field program for sound over a flat ground through a layered atmosphere, written apart from the engine: the exact solution of the vertical problem through layers of constant sound speed under a Hankel transform, with its own Bessel functions, so it needs nothing but numpy. In still air it is within 0.001 dB of the exact solution for a locally reacting ground from 10 Hz to 1 kHz and 100 m to 10 km, and its layers are converged.validation/benchmark_1995.py: the benchmark cases for outdoor sound propagation models (Attenborough and others, 1995): a source 5 m and a listener 1 m over a ground of flow resistivity 200 kPa s/m², 100 m to 10 km apart, in still air and in sound speed gradients of 0.1/s down and up, and the same gradients stopping 100 m up (the weather there is), at 10, 100 and 1000 Hz through the engine’s parabolic equation and, as whole scenes, per octave band from 63 Hz to 1 kHz, against the program read as the engine reads its bands. The parabolic equation is within 0.1 dB of the program to 1 km and within 3 dB to 10 km with the gradient stopping 100 m up, as the weather’s do, and parts from it beyond 1 to 2 km with the benchmark’s gradient going on to a kilometre up; whole scenes are within 0.11 dB per band to 1 km in every case, in still air read through the band filters within 0.04 dB to 200 m. The comparison found the notch at the join of the parabolic equation’s bands and the rays’ fixed below; see Validation.
Outdoor scenes on the GPU
The GPU traces outdoor scenes as the CPU does, so
simulation.late.device: gpu,bothandautotake them now, where they fell back to the CPU: the rays bent by wind and temperature follow the same chords through the layers (the same turns and the same curve to the millimetre, ray for ray, worked in single precision without trigonometry, whose rounding on a GPU would move the chords’ ends); the rays through vegetation meet the same foliage and stems, scatter and lose energy the same way, and carry the same per-band weights; the open sky lets them out (a surface into no room, which only the CPU took before); and the ray count is told from the chords it needs. Vegetation of more than 16 media stays on the CPU, which the run says. What reaches each listener in each band, and when, agrees with the CPU’s within the noise of the ray count:validation/outdoor_gpu.pycompares the two on the outdoor examples (see Validation). The shader is compiled with only what a scene has of the outdoors, so rooms compile as fast as before. On Windows, Direct3D 12’s shader compiler (FXC) compiles the GPU’s shaders without its optimizer, which got the vegetation wrong (the driver optimizes them anyway, and they compile as fast); there a scene with both bent rays and vegetation stays on the CPU, which the run says, as FXC takes more than an hour over the tracer with both.The wave solver in wind and temperature steps its grid on the GPU too (
simulation.wave.device: gpu,bothandauto), with the speed of sound per cell, in its walls, its source and its matched layer, agreeing with the CPU to rounding.
Eddies weaker with height
atmosphere.turbulence.height_msays the height above the ground the structure parameters C_T² and C_v² hold at (where they were measured or estimated, usually 2 m): from it the eddies weaken with height as (z /height_m)^-falloff, taken no nearer the ground than their outer scale.falloffis 2/3 by default, as the surface layer’s structure parameters fall in neutral and stable air (Monin-Obukhov similarity), up to 4/3 for the temperature’s in strong sunshine. Withoutheight_mthe eddies are the same at every height, as before: results do not change.The coherence of each pair of early paths takes Cn² at the mean height of the two points along them, so paths aloft part slower; the sound scattered into a shadow takes Cn² at the height it scatters at, and the shaking of the refraction near the ground Cn² at the geometric mean of the source’s and the listener’s heights.
validation/turbulence.py --profilemarches the parabolic equation through eddies strengthening right down to the ground: the ground dip comes out within 1.1 dB on average over the soft grounds and 1.4 over the hard ones, the shadows within 1.3 to 2.9 dB; the thin layer of stronger eddies at the ground does not tell there, which is why the engine stops at the outer scale (see Validation).inspectnames the height and the fall-off; the editor’s Wind and temperature group has Eddies’ height (0: the same at every height) and Weaker with height.
Grounds of extended reaction
impedance.reaction: extendedlets sound travel sideways inside a porous ground, so that each reflection meets the impedance of its own angle, Zc n / sqrt(n² - sin² θ) with n the ground’s complex index (and the layer’s with it), as Attenborough’s models have it. It tells for soft snow, moss, forest floors and thin layers: a layer of snow 10 cm deep on ice reflects up to 3.7 dB differently 20 m away, fresh snow 1.0 dB, grass 0.1 dB. The early paths take it at each reflection’s angle, within 0.4 dB of the exact field, the fast field program’s with the ground’s reflection at every angle (1.5 dB over the snow layer at 20 m, where the locally reacting impedance is off by up to 6.4); the parabolic equation takes it at grazing incidence, within 0.4 dB (1.1 over the layer). The default,local, is as before: results do not change. Not withtwo_parameter, which has no index; the wave solver’s ground reacts locally (see Validation).inspectnames the reaction; the editor’s Ground impedance group has Reaction.
Weather zones
atmosphere.zonesdraws parts of the scene with weather of their own, as polygons in the plan: the lee of buildings or a wood, where the wind drops; a lake, over which the air lies cooler under an inversion; a car park, warmer. A zone gives any of the atmosphere’s fields (preset,wind, the temperature’s gradient and log terms,roughness_m,table) andtemperature_c, the temperature at the ground there; what it leaves out it takes from the scene’s atmosphere. Outside the zones the scene’s weather holds, the first zone wins where two overlap, and the eddies are the scene’s everywhere. Without zones nothing changes: results are as before, bit for bit.The parabolic equation marches through the zones, changing its speed of sound where its path crosses a border, and the wave solver takes each cell’s zone; the early paths bend through the mean profile along their way, each zone’s weighted by the length of the way over it (the equivalent profile of Nord2000 and Harmonoise), and the late rays through the mean weather over the scene’s extent, each zone’s weighted by its area. Checked against a Crank-Nicholson parabolic equation marched through the zones (
validation/weather_zones.py): the engine’s parabolic equation is within 0.15 dB on average and 1.2 dB at most to 800 m over shelter and gusts downwind and upwind, an inversion past a lake shore and cool water downwind; the mean profile within 0.3 dB where only the temperature changes on the way, and 1 to 4 dB, 6.6 at most, where the wind changes (see Validation).inspectnames each zone, its corners and its weather. The editor’s Wind and temperature group has Weather of (the whole scene, or a zone), Add weather zone, which draws one in the plan, Remove weather zone and a zone’s Ground temperature; the plan draws the zones dotted in sky blue, with their names.
Levels over a long term
long_termgives weather classes and their shares of the time, and the scene is run in each: noise maps and assessments give the level over a year, in which each weather counts for the share of the time it holds (ISO 1996-2, CNOSSOS-EU, Nord2000). A class takes any of a weather zone’s fields over the scene’s weather and may have eddies of its own; itsshareis one number or one per period of the day. The periods are L_den’s by default (day 12 hours, evening 4 with 5 dB, night 8 with 10 dB);long_term.periodsgives others. Every run shares the seed, and a class whose weather is the scene’s takes the scene’s run. Withoutlong_termnothing changes: results are as before, bit for bit.At every listener and map point, per octave band and A-weighted: the level in each class, in each period (the energy mean of the classes by their shares), the long-term level (the energy mean of the periods by their hours), the rating level (L_den, or L_R with other periods), and the long-term level relative to free field and insertion loss.
simulateprints them,parameters.jsonhas them in along_termentry per response, maps gainlong_termandrating_level, andinspectlists the classes and periods. Checked byvalidation/long_term.py: the combination agrees with the formulas worked from each class’s run to rounding, and the correction from the classes favourable to propagation to the long-term level follows ISO 9613-2’s C_met within 0.8 dB, with a C_0 of 3.4 dB (see Validation).The editor’s Weather of lists weather classes as well as zones, with Add weather class, Remove weather class and each class’s Share of the time per period (the other classes make room in proportion); the results panel gains a Long-term level row and the run details each period’s and class’s levels and the rating level; maps offer both new parameters.
Moving sources and Doppler
A walk with
doppler: true(a car, a train, an aircraft, or a listener passing quickly) follows each early path from place to place and plays it through a delay that changes as the walker goes, read through a windowed sinc, so each path’s pitch shifts as its length changes; its filters change from place to place as before, in turbulence the eddies drift past and binaural takes the ears’ filters. The late part is crossfaded as before. A walk without it renders as before, bit for bit.Checked by
validation/doppler.pyagainst the exact field of a source passing over a rigid ground: within 45 dB at 30 m/s and 38 dB at 80 m/s, and the level within 0.42 dB at every moment, where crossfading misses it by up to 11 dB; over grass the sound keeps the energy of the places’ responses within 0.07 dB in every band (see Validation).Pass-by levels:
acoustic_spaces.walkthrough.pass_bygives the level at the listener over time (A-weighted and per octave band, the source standing at each place), its maximum, the sound exposure level and L_Aeq; the walkthrough of a walk with Doppler prints them,walkthrough --levels CSVsaves the levels of any walk, andinspectsays which walks have Doppler. Faster than sound or with the wave solver’s bands, a walk with Doppler stops with an error.
Moving air in the wave solver
simulation.wave.wind: moving(simulate --wave-wind moving, Wind on the editor’s Simulation page) makes the wave solver’s air move with the wind instead of changing its speed of sound: each cell keeps the still air’s speed of sound, and before each step the pressure is carried along with the wind (fourth-order differences, a third-order Runge-Kutta method). Sound goes at c + v downwind and c - v upwind whichever way it turned, so a façade’s echo against the wind comes when it should. It costs three to four times as much, on the CPU (asked for the GPU, the run says so and stays on the CPU); the run’s details saywave_moving_air. Left out (effective, the default), and in still air, nothing changes.Checked by
validation/wave_wind.pyagainst the exact field in a uniform wind of 15 m/s: within 0.05 dB and 0.1 ms in every direction, where the effective speed of sound is up to 0.9 dB off, and a façade’s echo within 0.02 ms of ray theory’s, where the effective speed of sound brings it 10.6 ms early; in thewindypreset over grass within 0.34 dB of the parabolic equation, 0.13 dB on average, where the effective speed of sound is within 0.69 dB (see Validation).
Whole cities
Edge diffraction skips the pairs of an edge and a plane that no path reflecting between them can use, before it traces any ray: the edges behind the plane, and those whose shadow, cast on the plane from the source’s or the listener’s image, misses the plane’s faces. For the listener’s images it also skips the paths that would touch the edge’s line beyond its ends or arrive too late. The exact tests run on the rest as before, in native code on
simulation.performance.threadsthreads, so the results are the same bit for bit. A city of 900 buildings simulates in 2.6 s on four cores instead of 81 s and 7.1 GB, 1,600 buildings, which ran out of 14 GB, in 4.8 s, and 3,600 in 24 s; with the reflections that fade at edges (below) in 17, 21 and 56 s. The work still grows with the square of the buildings, as about 3 % of the pairs can carry a path as far as geometry alone tells.validation/city_speed.pytimes cities of growing size and checks the results against testing every pair (see Validation).
Wide-angle parabolic equation
simulation.early.parabolic_equation: wide_angle(simulateandmap--parabolic-equation wide_angle, Parabolic equation on the editor’s Simulation page, for an outdoor scene in wind or temperature gradients) gives the bands the parabolic equation gives outdoors (up to 500 Hz, and in a shadow up to 2 kHz) from a wide-angle parabolic equation,acoustic_spaces.engine.wide_angle: the split-step Padé method of Collins’ RAM, the one-way operator as a rational function of ten terms in steps of three wavelengths, fourth-order differences in height with the ground’s impedance as the boundary, and the operator’s own starting field, its tridiagonal solves in native code. It carries sound that a gradient reaching hundreds of metres up bends back at up to 45 degrees, through air whose speed of sound is far from the ground’s, where the narrow-angle equation strays over a few kilometres. It takes three to four times the work; left out (narrow_angle, the default), nothing changes.Checked by
validation/benchmark_1995.pyagainst the fast field program: within 0.11 dB at the benchmark’s pure tones from 10 Hz to 1 kHz and 100 m to 10 km wherever the sound is within 40 dB of free field, where the narrow-angle equation strays by 3 to 20 dB beyond 1 to 2 km; in the weather presets within 0.04 dB on average at the bands’ centres and 0.02 dB over the bands (the narrow-angle one 0.11 and 0.08 dB); whole scenes of the benchmark’s cases 2 and 3 within 0.03 dB per band to 2 km, where the narrow-angle equation is 0.5 to 0.8 dB off at 2 km downwind. Through weather zonesvalidation/weather_zones.pyfinds it within 0.09 dB of the Crank-Nicholson reference on average (see Validation).
Coherent arrays element by element
near_field: trueon a source’s array (directivity: {file: line_array.yaml, near_field: true}) adds up its elements’ sound one by one wherever an early path leaves it, each from where it is, with its own distance, direction, feed and pattern, instead of sending out the array’s far-field pattern from one point. Listeners nearer than about the array’s length squared over the wavelength hear its sound as it is there: on axis a 3.75 m column loses about 3 dB per doubling of the distance up to 41 m at 1 kHz, as a line does, and 6 dB beyond. The direct sound, the reflections, the ground’s reflection and refracted rays outdoors, and the sound that reaches an edge all take it; the late reverberation keeps the pattern. The room no longer warns about such an array’s near listeners, and the warning for the others names the option. In the editor: “Add up the array’s elements one by one” under the source’s directivity. Without it nothing changes. See Validation.
Groups, and sources left out
Sources and listeners take a
group(nested with/, asPA/Delays) andenabled: false, which leaves them out of simulations, maps, walks’ other sources and the scene’s checks while keeping them in the scene; the others get the same responses as in a run of all. Named with--sourceor--listener, they are simulated all the same;inspectmarks them “left out”. See Groups, and sources left out.The editor’s scene tree shows sources and listeners in their groups. A tick on each source, listener and group uses it in simulations or leaves it out, in one step of the history; an eye hides it in the 3D view and the plan or shows it again, without changing the scene. Sources and listeners left out are drawn faded. Their form has a Group field and Use in simulations.
The
stadium_pa,concert_stageandstadium_concertexamples put their sources and listeners in groups: the PA’s mains, subwoofers, fills and delays, the monitors, the stage, and the audience standing and seated.
Examples
New examples of a concert’s sound system, made of the same stage, arrays and boxes:
stadium_pa, a football stadium’s sound system as tours hang it: two main line arrays of 12 boxes in a J, a block of eight cardioid subwoofers timed to arrive with them at the mix position, out-fills for the side stands, two in-fill arrays under the stage roof’s front edge for the pitch’s centre from 10 to 30 m, where the main arrays’ horizontal coverage ends, front fills, two delay towers on the pitch and four delays under the roof’s edge, each fill and delay arriving 6 to 20 ms after the main arrays at its seats (the in-fills with them), tuned at the mix position as a system engineer tunes it (the main arrays’ low mids 1 to 2 dB down, their top octave 4 dB up for the air, the subwoofers 2 dB under them, the fills’ low end rolled off). The stage’s back is hung with drapes, as touring stages are, and the late sound reaches the listeners as secondary sources: in a bowl this size a detector sphere catches so few rays that a single one would be heard as an echo. Listeners stand on the pitch and on the stands. The crowds standing and seated scatter as a measured theatre audience does, rising from 0.3 at 125 Hz to 0.7 from 1 kHz (scattering: theatre_audience) rather than at one value; the rows stay out of the model, as surfaces smaller than the wavelength would only mirror sound falsely.align.pyprints when and how loud each loudspeaker arrives at each listener, and which would be heard as an echo (none).concert_stage, the stage alone on open ground with its monitors (wedges for the singer, a drum fill and side fills), the main arrays, the subwoofers and the front fills; the players and the mix position listen. Each player hears their own monitor loudest.feedback.pyfinds how much gain the singer’s microphone allows before the loudspeakers make it ring, in each octave band at the mix position and at the singer’s ears, the frequencies where it rings first, and renders the loop below and above that gain.stadium_concert, the two together.
Every array adds up its boxes one by one (
near_field: true);near_field.pyprints how far the arrays’ far-field patterns would be off at each listener (within 3.5 dB up to 2 kHz, up to 10 dB at 4 and 8 kHz), skipping the bands a source’s filter cuts.make_model.py, the same in the three, writes each one’s model and arrays, of a generic box made up for the examples.The editor’s File > Open example shows the examples in groups (rooms, halls and churches, sound systems, outdoors, models from CAD and BIM files, measurements and material libraries), and Example scenes lists them in the same groups;
acoustic_spaces.examples.example_groups()gives them by group.
Demos
Two more demos with their scripts, beside the orchestra in the concert hall:
church_demo.py, in thechurchexample, puts a choir of 19 real singers (VocalSet’s dry takes, CC BY 4.0, brought onto the beat and towards the tune) and a sampled pipe organ (VCSL, CC0) in the church;stadium_demo.py, instadium_concert, adds a band to the stage and plays a funk song through it, the sound system and the monitors, from the song’s own stems (“Spicy Funk Cake”, recorded for the DEGA, CC BY 4.0): the drum kit plays the drums, the singer the vocal, the guitarist’s amplifier the guitar; its video turns each stop up to the loudness of the loudest, the stage, and says by how much, and its pictures show the crowd standing on the pitch and seated in the stands, the empty stand’s seats, how many people stand and sit, and how much of the sound each takes up.--song franciumorrubidiumplays one of Software-Entwicklungskit’s electronic songs (CC0) instead, from the stems its album is released with, all twelve listeners in the tour. Each downloads its recordings, writes its scene, mixes the piece at every listener with one gain for all and makes a video that tours them (with ffmpeg and Pillow).orchestra_demo.py, inorchestra, tours the concert hall’s four seats through the Beethoven recording’s three movements in one video: the seat changes at the quietest moment near each share of a movement, the movements follow each other with a pause, every seat’s mix is shifted earlier by the time the sound takes to reach it, and all stops keep one gain, so the hall’s differences are heard as they are; the cover says each stop’s level against the loudest and carries a bar with the tour.lakeside_demo.py, inlakeside_concert, plays the same song from the lakeside stage through a sound system as a lawn concert hangs it (two line arrays of eight boxes and four cardioid subwoofers, aimed at the lawn and tuned at the audience) and mixes it in the audience, behind the stage, on a boat and in the village across the lake, in the evening’s weather and in still air; its video plays the song at a bedroom window in the evening’s weather, in still air and in the evening’s again, each stop turned up to the audience’s loudness with its real level in the caption, over a slice across the lake with the sound’s rays bent by the weather.Demos gathers the demos, with where each recording comes from, its licence and the credit a shared render needs; the orchestra’s section moved there from spatial audio and listening. The README and that page link the playlist of the videos.
Every demo script takes
--hrtf me.sofa, a SOFA file of your own or measured HRTFs: thesimulatecommand it prints carries it, andrendermixes from the responses made through that file, refusing ones made through the spherical head. A script asked for feeds or mixes before its scene exists says to run its scene step first.A fifth demo, and the first of speech:
walk_demo.py, in thegalleryexample (into whichgallery_walkis merged, see Movement), has a guide talk while a visitor walks away from them down an L-shaped gallery and round the corner, where only the sound bent round the corner arrives (56.0 dB(A) and a speech transmission index of 0.73 at the start, 48.7 dB(A) and 0.54 round the corner), and then stands in front of them and turns their head a quarter at a time through the engine’s head tracker. The talk is the ACE corpus’s anechoic speech (CC BY 4.0); the video follows the visitor from above.The demo scripts share one package,
acoustic_spaces.demos: a demo describes its steps and its options, and gets the command line, the recordings and their phrases, the simulate command and its HRTF, the mixes with one gain, the crossfaded tour, the stills and the video from the package.python NAME_demo.py allruns a demo from nothing to its video, leaving out the steps whose result is there already (the simulation among them, now a step like the others),--againmixes the stops afresh, and the same demo runs from the editor. Roughly 2000 lines of copied code went; what the demos do is unchanged, except that the church’s tour now shifts each stop by the time the sound takes to reach the listener, as the other tours did.acoustic_spaces.fetchholds the downloads the demos and the HRTF sets share: a whole file, or one member of a remote zip archive read where it is by HTTP range requests.Any scene makes a demo of its own:
acoustic-spaces demo scene.yamlplays the WAV file each source’s newsignalnames, mixed down to mono and brought to one gated RMS so that the sources’ sound powers set the balance, as simulated; sources without one stay silent. It simulates the scene, mixes every listener in use with one gain for all and makes a video that hears each in turn for--seconds(or the listeners--tournames), its stills drawn from the scene’s own outline from above, with the level, the arrival and the reverberation time there. It writes intodemo/beside the scene (or--work), binaural through the HRTFs the scene’s binaural output names (a spherical head model when it names none) unless--hrtfsays otherwise.acoustic_spaces.demos.scene_demois the demo, the stepsfeeds,simulate,renderandvideo.Every demo’s stills say how the sound was made: the version, the early reflections (traced or image sources, their order, diffraction) and the late sound with its rays, as the simulation ran them, the wave solver’s bands when it ran, outdoors in moving air the bent paths and the parabolic equation’s bands, the HRTFs the mixes went through (or the spherical head model) and the sample rate, and the command that makes the video again with the options it was given, paths by their names (the song, the HRTFs).
A demo run elsewhere than its example (the editor’s copy under the home folder, when the examples cannot be written) writes scenes that find the example’s model and data: their paths were left relative to the example and the simulation could not find them.
Every demo takes
--input SOURCE=WAV, once for each source: the source plays that file instead of what the demo gives it, mixed down to mono, at the gated RMS of the feed it replaces and cut or padded with silence to its length, so the balance and the tour hold (the walk plays the guide’s file as it is; the demo of any scene feeds it as it would the signal). The mixes are made again when the inputs change, and the stills’ credit line names the files. Before, only the demo of any scene could play other recordings, by changing the scene.The stadium and the lakeside demos play a song of the user’s own in place of their presets:
--music FILE(a stereo master: WAV, or anything ffmpeg decodes) with--bpm, its tempo, which the tour’s phrases of 8 bars go by;--creditsets the line the video shows (the file’s name unless given). The stadium also takes--stem NAME=WAVfor each stem (drums, bass, guitar, keys, brass, vocal, backing), which the monitors’ mixes and the band play, and--downbeatfor the first downbeat; a mix none of whose stems is given plays nothing, andfeedsand the Run demo window’s table say so. Nothing is downloaded, and the music goes to a folder named after the file. In the Run demo window,--musicis a file field,--bpma number and the stems a table with a file per row; the window’s Demo list is left out when there is one demo only.
A car passing by, with Doppler
The
roadsideandstreetexamples have a car: a point source 0.05 m above the road with the sound power per octave band of CNOSSOS-EU’s passenger car (category 1) at 50 km/h, and a walk along the road with Doppler. It is left out of the examples’ own simulations (enabled: false) until ticked, so their results do not change.passby_demo.pyin both examples drives it past every listener in turn: the car’s sound made to CNOSSOS-EU’s sound power at its speed, the speed of its walk in the scene or--speedin km/h (tyre noise and the firing of a four-cylinder engine at--rpm, each with the octave spectrum of its part, the two at the ratio of their powers;acoustic_spaces.demos.car), or a dry recording of yours (--input car=WAV); the car simulated every--step(1 m) along its walk at each listener and its sound played through those places, each early path through its own changing delay; and a video with the car moving on the plan, the level heard over the pass under it, L_Amax, the sound exposure level and how far the pitch of the direct sound falls. One module,acoustic_spaces.demos.passby, serves both. The car’s walk goes through points at a speed or through keyframes timed for one steady speed (within 1 %), as CNOSSOS-EU gives a car’s sound at a steady speed;--speeddrives it through the keyframes’ places at the speed given. In the Run demo window--speedshows “km/h” and starts at the walk’s own speed (“as in the scene”), each number shows its unit, and a line under the options gives the walk’s length, the speed in km/h and m/s and how long the pass lasts, as they change (Option.unit,Option.unsetandDemo.summaryfor any demo).The demo of any scene plays a walk with Doppler when it is the only walk that moves: every listener hears it pass, with the moving picture and the level at each place, all from the one simulation of the moving run (
render_movement). With other walkers it says to leave them standing.Editor: the Movement panel has a Doppler tick (
walk.doppler), one step of the history; a line or a surface source’s form shows Moves and Edit movement… greyed, with a tip saying that only a point source moves, where it hid them before. The Run demo window’s Movement box lists a walker left out of the scene when the demo moves it, as the pass-by demos do the car.
Movement
A walk may go through
keyframes: where the walker is at given times after its start and, if given, how it faces there, between which it moves and turns evenly (the shorter way round), so it can wait, turn on the spot or go at any pace. A keyframe without an orientation keeps the facing of the one before.walkthroughplaces such a walk every--stepalong a leg and every--turn-step(20° by default) of a turn on the spot.walk.enabled: falsekeeps a walk but leaves its walker standing still.walkthrough --allplays every walk that is enabled at once, on one clock, in one simulation: each mover as a copy at each place it takes, with the still sources and listeners, the late part traced every--late-stepalong each walk; every listener in use hears every source, and one WAV is written per listener (or--listener’s), with one gain for all.--videodraws them all going. A walk with Doppler moves alone: it plays with Doppler when it is the only walk that is on, and with others on as well the run says to leave them standing (Movement.doppler). From Python:walkthrough.knots,movers,simulate_movement,render_movement,save_movementandviz.animate_movement.inspectdescribes a walk along keyframes and one that is off.Editor: Edit > Edit movement… (or the button in a source’s or listener’s form) opens the Movement panel: the keyframes of the selected point source or listener in a table (time in s, position in m, angles in degrees, with the speed in m/s to the next keyframe shown, not set), edited in place as one step of the history each; Add keyframe at the playhead, Remove keyframe, Convert to keyframes for a walk along
to; a walk alongtoedited in place too, its start and points in the table with the time each is reached, a point typed or dragged, Add point and Remove point, and one speed for the whole walk typed in its Speed column; a timeline whose End can be made longer to add keyframes past the longest walk, with Play and a slider that move every mover in the 3D view and the plan; walks drawn as dashed lines, the selected mover’s keyframes as numbered handles to drag in either view; Moves (also in the forms) forenabled; and Render movement…,walkthrough --allin the background, writing one WAV per listener with one gain, in the panel’s Output: binaural first (through a spherical head when the scene has no binaural output), then the scene’s own outputs; a walk with Doppler plays with Doppler. Before, it wrote the scene’s first output, mono in theroadsideandstreetexamples, so the car passed no ear, and a walk with Doppler was refused.One scene serves still and moving:
gallery_walkis merged intogallery, whose visitor walks down the long arm and round the corner unless its walk is turned off, beside the two listeners that stand. The namegallery_walkstill finds it (examples,examples --copy, which copies it under that name, and the demos).walkthrough.with_movingturns the walks it names on and every other one off, for a run, without changing the scene.The demo of any scene plays the walks that are on: every walker moves on one clock from the start of the recordings, the mixes come from one simulation with every walker at each moment of its walk (in
moving/, one gain for all), and the video is a moving picture with the walks dotted, every walker where it is and the level of the sources that play at that moment, as simulated.--moving NAME,...picks the walks of a run and--moving nonekeeps every walker standing; a walk with Doppler moves alone (above).The Run demo window’s Movement box lists the sources and listeners with a walk, how each goes, how far and for how long, ticked when it moves in the run: This scene’s ticks start as Moves sets them and change that run only (
--moving), As in the scene ticks them as the scene does, an example’s demo shows what it moves, and Edit movement… closes the window and opens the Movement panel on the walker.walk_demo.pywalks the visitor as the scene has it, along keyframes too, with the pace its captions state worked out from the walk; it makes its scene and its walk again when the example’s scene has changed since, and says when the visitor’s walk is off.
HRTF sets
acoustic-spaces hrtflists free HRTF sets with their licences and downloads the one named into the user’s data folder:ku100(a Neumann KU100 dummy head, TH Köln, CC BY 3.0, 3.4 MB),kemar(a KEMAR dummy head, SADIE II, Apache 2.0, 38 MB) andviking(a KEMAR dummy head with one of 20 pairs of replica pinnae or none, Viking v2, CC BY 4.0, 58.4 MB, one file per pair of pinnae with--member).--urltakes any other SOFA file,--folderkeeps them elsewhere, and the licence and who measured the set are printed before anything is downloaded. Until now a measured head had to be found and fetched by hand.
Running on another computer
acoustic-spaces pack scene.yaml scene.zip(File > Pack for another computer in the editor,acoustic_spaces.pack) writes a scene and every file it names (model, terrain, buildings map, material and directivity files, the WAV files sources play, HRTFs, a glTF model’s buffers, an array’s element data) to one zip, with the scene file naming them relatively. Every command that reads a scene (simulate,map,validate,calibrate,walkthrough…) takes a pack in its place and gives the same results, bit for bit. Itspack.jsonrecords the version that wrote it and a SHA-256 of every file; damaged packs are refused and packs from another version warn. Packing uploads nothing. Scenes with problems are not packed.A
Dockerfilebuilds an image with the command line, headless, on the GPU through Vulkan: Mesa’s drivers inside, and NVIDIA’s from the host with the driver’s graphics capability, which the image asks for.docs/remote.mdruns packs with it on a server over SSH and on Hugging Face Jobs. CI builds the image and runs packs in it on the software Vulkan driver.Scene.yaml_text()writes the scene file’s text with the files it names written as a function says.
Editor
The wind is drawn outdoors (View > Show wind): sky blue arrows over the ground in the 3D view, in
viewand in the plan, the way it blows and as long as it is strong, at the height the scene gives its speed at, clear of the sources and listeners; each weather zone with its name and wind (and its ground temperature when it has its own), in the 3D view as see-through walls; and the scene’s wind in a corner. View > Show wind profile stands a mast in the 3D view in each weather where the wind blows, with arrows up it and a curve: how the wind grows with height over the ground. The scene tree’s new Wind entry has an eye for each and lists the weather zones, which show their weather in the scene’s form when selected. Before, only the Show the profile plot showed the wind, and nothing showed a weather zone’s.File > Export particle animation to Blender asks how much of the sound to animate, how slowly and how many particles, as
export --duration --slow-motion --countdo; outdoors it starts with sound enough to reach the farthest listener, slowed down to last about the default’s 4 s. Before, it always took 100 ms, too little to leave the stage outdoors.New example:
lakeside_concert, a concert on a lawn by a lake heard across the water from a boat and in a village 550 m away, with two weather zones: cool air and a stronger breeze over the lake, a weaker breeze among the village’s houses. The editor draws its three weathers with Show wind and Show wind profile.While a run, a coverage map, a tuning or an export goes on, the status bar says what runs, its latest step and the time so far, with a Cancel button beside the bar; resting the mouse on the bar says the same at length. The bar was small and silent before, easy to miss on some desktops (KDE’s, say).
Messages shades every other message, and its ticks show or hide the info, the warnings and the errors, with how many there are of each. Item keeps only the messages about one source, listener or material, a filter only those holding some words, and Clear removes the messages on show.
The markers of sources and listeners in the 3D view and
showtake their size from the scene: a fifth of the usual spacing between the points, from 1.2 % to 2.5 % of the model’s size (it was 1.2 % always), so in most rooms they are twice as large. Where they would overlap they get smaller, down to a fifth, and an arrow that would run into a marker ahead of it stops short, in the plan too: on the stadium’s stage the markers were 6 m wide and the arrows 19 m long, across the players and monitors. Inshow, sources and listeners left out are drawn faded too.Simulation > Run demo… runs a demo of the open scene: This scene, with the WAV files its sources play, and for an example’s scene (or a copy of one) the example’s own demo, reading the open scene. Pick the demo and the step (all for the lot), fill in the options, which start at the defaults the command line uses, and it runs in its own process in the background, with what it prints in Messages and Cancel to stop it. The window shows the command line it is about to run, and asks for a scene with changes to be saved first. A demo could only be run from a terminal before.
The Run demo… window lists the sources of the demo in a table, what each plays and the WAV file it plays instead (
--input): Choose WAV… gives the selected source one, Default takes it back.The source page’s Plays row names the WAV file the source plays in a demo of the scene: Choose WAV… picks one, Silent takes it away, and the scene file keeps it as
signal, relative to the scene’s folder.Download HRTFs…, on the Outputs page and beside the Run demo… window’s HRTF field, lists the free HRTF sets with their licences, downloads the one chosen in the background and puts it in the scene’s binaural output; another SOFA file goes in by its address.
Nothing starts cut off. Properties takes the window’s whole height on the right and is never narrower than its widest form needs (up to half the screen): the window opens, and stays, wide enough for it beside the views and the panels under them, so a larger font or a longer language widens the window rather than cutting off the forms. A form scrolls by its own height, not the tallest form’s, and a row too long for the panel puts its field under its label. The Run demo… window opens wide enough for its paths and scrolls where the screen is too short, rather than squeezing its tables and buttons; Download HRTFs… opens as wide as its rows. Every window opens within the screen it is on. The ground’s Extent takes two rows, one per corner.
A copy of an example saved anywhere keeps the example’s demo in the Run demo… window: the example is known by the folder’s name or, failing that, by the scene’s name (
examples.example_of).
Reflections that fade at edges
Every specular path fades in and out at the edges that cut it off, instead of coming or going at full strength: where a leg of it passes an edge, and where the point it reflects at passes the edge of its face (
engine.edge_fade). The edge diffraction bends the direct sound and the first-order reflections round edges; a path with more reflections, such as the copy of a first-order reflection off the ground or a second-order reflection, stepped at its boundary: in thestreetexample’s garden the passing car’s early level stepped by up to 11.4 dB(A) from one metre of road to the next. Every path, found or not, now takes the uniform theory of diffraction’s term of its own boundary at each edge near each of its events. A path that is not found is followed where one edge explains each run of the events that stop it, or two edges explain them between them, in at most two runs, as a reflection past the end of its face with a corner in the way of a leg next to it. The terms of one event add (each edge on its own side, as round a gap), the events of a path multiply, and the terms the edge diffraction adds already are left out. Reflections off perpendicular planes (the floor and a wall) give one image in either order: each path takes the terms of every such order, and tries each where the early stage found only one. Each term is a path over its edge, and each product of terms a path of its own, which arrives later by the detours of all its terms. It comes withdiffractionon (any mode butnone).Against the exact solution (
validation/edge_fade.py: BTM with images in a rigid ground, listeners 1 cm apart past a screen’s end, a building’s corner, a strip 2 m wide and a gap 4 m wide): within 0.08 dB in every band (0.19 dB at 8 kHz on the strip, with or without the fade), and no step larger than the exact solution’s own; without the fade up to 7 dB off, with steps of up to 9.2 dB.Round the
gallery’s corner, walked 1 cm at a time with reflections up to third order, the level steps by at most 0.13 dB in any band, instead of up to 3.8 dB (see Validation, which lists what is left). Where the edge diffraction bends every path the early stage follows (max_order: 0withdirect,max_order: 1withfirst_order), the fade changes the level by less than 0.05 dB.Driving the
street’s car 1 cm at a time, the garden’s level no longer steps by up to 1.9 dB in one band (0.63 dB(A)) where a path that two events stop clears one of them: at those places it now stands out from its neighbours’ by at most 0.11 dB.Indoors, in the bands where paths add in energy (from 250 Hz up), the terms of one event take their cross terms with each other and with those the edge diffraction adds there, and the edge diffraction’s waves of two edges that border one path take theirs, as one wave: round a narrow face the two edges’ terms nearly cancel, as they do in pressure. Behind the
archway’s wall the level stepped by up to 4.4 dB at 250 Hz to 1 kHz where the reflection off the arch’s jamb appears, now by up to 1.4 dB. The early energy there changes by -2.2 dB at 250 Hz to +3.5 dB at 2 kHz: nearer the sum with every band in pressure from 1 kHz up (at 2 kHz 0.3 dB below it instead of 3.8 dB), further at 250 Hz (1.6 dB below instead of 0.6 dB above); through the arch it changes by up to 0.4 dB. In the hall with alcoves ofvalidation/edge_diffraction.py, withfirst_order, it is 1.8 dB below that sum at 250 Hz instead of 5.6 dB. Theleaky_classroomchanges by 0.13 dB, the other examples by less than 0.01 dB.Known issues (Validation): steps are left at the corners of edges, where the point a term diffracts at leaves the end of its edge: driving the
street’s car past the barrier 1 cm at a time, up to 1.42 dB in one band and 0.35 dB(A) where the paths over its top reach its ends. Indoors, where paths add in energy from 250 Hz up, behind thearchway’s wall the level steps by up to 1.4 dB at 250 and 500 Hz where the reflection off the arch’s jamb, 40 cm deep, appears (2.2 dB without the fade). Not checked against an exact solution: paths of two or more reflections, surfaces with an impedance,second_orderandboth_sides; nothing against measurements.Results change wherever paths with more reflections than the edge diffraction bends pass an edge. The examples’ early energy changes by up to 5.2 dB at 63 Hz behind the
archway’s wall (4.2 dB at 125 Hz), 2.8 dB at 63 Hz in theleaky_classroom’s back row, 1.4 dB at 250 Hz in thestreet’s garden, 1.4 dB round thegallery’s corner, 1.0 dB in theconcert_hall’s rear stalls and 2.2 dB at 63 Hz in theorchestra’s middle stalls (a cello’s second- and third-order reflections that pass just over the front edge of the stage, as its direct sound does), over every source and listener; the other examples are unchanged (most have no edge that diffracts, or diffraction off).Each term is a path, so the early stage costs more where edges diffract: in a hall with 16 edges 296 ms per source and listener instead of 83 ms (981 paths instead of 71), the gallery’s coverage map 34.3 s instead of 16.8 s, and a city of 900 buildings 16.6 s instead of 2.6 s.
BVH queries of fewer than 256 rays run on one thread, and larger ones on a thread per 256 rays up to the cores: starting a thread for every core took longer than the rays (100 rays took 0.14 ms, now 0.02 ms). The results are the same bit for bit.
Checks of the decay times (ISO 3382-2, annex B)
Each T20 and T30 comes with the non-linearity ξ of the line it is read from, 1000 (1 - r²) ‰ (r: the correlation of the decay curve with time over the fit range), and the decay’s curvature, 100 (T30 / T20 - 1) %:
xi_t20,xi_t30andcurvatureinRoomParameters, the printed tables,parameters.jsonand the editor’s results table.NON_LINEARITY_LIMIT(10 ‰) andCURVATURE_LIMIT(10 %) are the standard’s limits: past them the editor’s results table shows the check in amber, a ξ with the T20 or T30 it puts in doubt, each with a tip that says why. A decay time left out for the background noise leaves its checks out as well.
Fixes
A response that ends in silence (all its sound arrived before its end, as outdoors) had decay times fitted through the drop after its last sound: the decay curve fell there to the floor of the computation (-3000 dB), and a fit that reached it took the drop for a decay. One reflection 1.5 s after the direct sound, then silence, read EDT 7.5 s. The drop is now the end of the data: a decay time needs the curve to fall through its range before the last sound, and is NaN otherwise. The same holds where removing a measurement’s noise leaves no energy. The energies of such a response (C50, C80, D50, Ts, G, lateral fractions, stage support) do not change.
Binaural renders lost treble: at 8 kHz up to 13.6 dB from some directions (the spherical head’s ear facing a source straight ahead), 2.7 dB on average over all directions, and more the higher the band. An HRTF between the measured directions mixes the responses round it aligned at their onsets, and an onset was the first sample within 20 dB of the peak: a response delayed by a fraction of a sample rings before its peak, more the nearer the fraction is to a half, so neighbouring onsets came out a sample or two apart and the mix cancelled the high frequencies. The onset is now the lag at which a response best matches its minimum-phase version (the same magnitude, starting at once), found to a fraction of a sample: the mix keeps the measured responses’ spectrum within 1 dB. Mono, Ambisonics and loudspeaker outputs were not affected. Every binaural output changes; demo videos made before this sound duller than they should.
A coverage map in the 3D view (and in
show) hid the sources and listeners: they stand at listening height, half under its tiles. The map is now see-through.A reflection right where the sound came from, on a surface over about 100 m across, printed an overflow warning (an error under
-W error). The results were right and stay the same.auralizeandwalkthrough(andrender.auralizeandwalkthrough.render_walk) mix the sources at their sound power. The impulse responses carry each source’s feed (gain, equalizer, delay) but are per unit of sound power, and the mix played every source as if it had the samelevel_db: in a sound system with fills 18 dB below its main loudspeakers, the fills came out 18 dB too loud. Each signal is now scaled by its source’slevel_dbagainst the loudest one’s in the mix;--equal-power(equal_power=True) keeps the old mix. A mix of sources of one power, or of one source, is unchanged.parameters.jsonlists each source’s sound power (sources), whichauralizereads; results written before it are mixed at one power, with a note.Levels, parameters and maps of open scenes take about half the time: the band filters that read the coherent bands were computed again for every band of every source and listener, and are now kept. The results are the same to the last bit.
In wind and temperature gradients the impulse response had a notch where the parabolic equation’s bands met the rays’ (around 700 Hz, up to 3 dB deep; around 2.8 kHz in the penumbra of a shadow), and the 500 Hz band of the level relative to free field came out 0.5 to 1.2 dB too low: the join between the two blended their spectra, which stand for different sound (the whole field over the ground on one path, each ray’s own on the others), and the silent paths’ higher bands counted against the band’s reading. Each band of the response now keeps its own spectrum into the crossovers, and a path that carries nothing in a band is left out of that band’s reading. The bands the parabolic equation gives now read back as computed, and the response follows the fast field program across the join within 0.3 dB. The same join in still air changes outdoor responses by less than 0.01 dB (one cell of the barrier tables by 0.2 dB, the ground dip at 250 Hz 200 m out) and rooms by under 0.0001 dB (the crossovers between the lowest bands’ filters are drawn band by band, a difference of a hundred-thousandth of the peak).
Saving an example elsewhere from the editor failed for the stadium (“directivity file not found: …/box.csv”): an array’s YAML file names its element’s pattern beside itself, and only the array file was copied. The files an array file names now go along with it (or, when one lies outside the example’s folder, the scene keeps using the original array file), and a save that fails leaves no copies behind. Saving the hillside example also took its terrain and map of buildings along but kept pointing at the originals; it now uses the copies.
Scene.files()lists every file a scene uses, the ones its array files name included. Saving also kept the window busy for seconds (Not Responding), adding up the arrays again from the copies; the copies hold the same bytes, so what was loaded stays: the stadium now saves in under a tenth of a second.The editor stopped answering while it exported a big run’s results (Not Responding for over a minute with the stadium’s 204 responses): the export ran in the window’s own thread. It now runs in the background, with progress in the messages and Cancel, and
write_results(and sosimulate --out) renders and reads the responses onsimulation.performance.threads(every core by default): the stadium’s export takes 47 s instead of 81 s on 4 cores. The files and the report are the same to the last bit.Wheels left out the hillside example’s terrain (
.asc) and map of buildings (.geojson), so the installed example could not load. A test now checks that every file the examples use is installed.second_orderdiffraction printed a floating-point overflow warning where two edges meet at a corner and the shortest path over both runs through it. The path is still turned down, as it was, and the warning is gone.On a hardware GPU the wave solver could stop partway through a long run on a big grid: it sent up to 512 steps in one submission, and a GPU whose work runs longer than its watchdog allows (10 s on Linux’s amdgpu, reached on an AMD Radeon 680M with two million cells) is reset, losing every step after it and failing the GPU runs that follow. The steps now go in submissions of about a second’s work each, which step the grid exactly as one long submission did.
The native code no longer fuses multiplies and adds that the source keeps apart, as Apple’s compiler did by default on Apple silicon: results there now round as on the other platforms, last bits included.
On Windows with Python 3.10 every GPU run printed four warnings that Vulkan lacked an extension (
Unable to find extension: VK_KHR_surface…): wgpu 0.31, the last for Python 3.10, warns of each one on a computer without a Vulkan driver, where Direct3D 12 does the work. They said nothing to a user and are left out now.
Known issues
A review of the code found issues that 4.0.0 still has, listed with what each one does in Validation and as open items on the roadmap: among them, no early path through more than two partitions while the late stage still leaves them out, G against a measured free-field response up to 3 dB high at 63 Hz when that response starts at its first sample, a source or listener inside a closed solid taken as a room, an earlier pack deleted when writing one of the same name fails, and two sources or listeners whose names differ only in characters a file name leaves out writing the same result file.
The README said “Status: 3.2”; it says 4.0 now and links the known issues.
3.2.3¶
Languages
Translations for Catalan and Italian for editor and command line text.
German and Dutch translations with the same coverage: editor menus, forms, tips and messages, plots, examples’ titles and descriptions, and the command line’s language report. Select Deutsch or Nederlands under View > Language (from the next start), set
ACOUSTIC_SPACES_LANG=deornl, or follow the system language, including the Windows and macOS display language.
3.2.2¶
Languages (acoustic_spaces.i18n)
The editor and the command line can show their text in another language, English or Spanish to start with. They follow the computer’s language; View > Language in the editor picks one for the user (from the next start), and
ACOUSTIC_SPACES_LANG=esorensets it over both.acoustic-spaces infosays which is in use and what chose it. Qt’s own dialogs and buttons follow it too. Scene files, commands, options and the files the commands write stay in English.The editor is in Spanish: menus, panels, forms, tips, messages and the tuning window, the choices in its lists (methods, reception, spectra, shapes, directivities; the scene file keeps its values), the coverage map’s parameters, the examples’ titles and descriptions, and the plots’ titles, axes and legends. Its numbers follow the computer’s regional settings (0,85 in Spain, 0.85 in Mexico), in what it shows and in what it reads from its tables and fields.
The Spanish text comes in steps: the command line and the library’s messages (the room’s checks and a run’s progress, which the editor shows too) come next, and until then show in English.
For contributors: text is marked with
_()and collected intolocale/*.pobytools/catalog.py, which CI runs to check that every catalog is complete; terms and style are in the new Translating page.
Fixes
The editor no longer aborts with a wgpu panic (
egl.rs ... BadAccess) when it looks for a GPU, as it does on opening the Simulation settings: with thegpuextra installed on Linux, wgpu also tried its OpenGL backend, whose set-up through EGL fails while the 3D view holds an OpenGL context, and that failure ends the process. wgpu now uses Vulkan, Metal or Direct3D 12 only, as the documentation says.
3.2.1¶
Wheels for Intel Macs
Wheels for Intel Macs (macOS x86-64), on PyPI and the GitHub releases, so pip no longer builds the package from source there. They need macOS 10.15 or later. CI builds them on GitHub’s Intel macOS runner and runs the tests and the editor’s tests there too; pull requests now build the wheels for CPython 3.10 as well as 3.14.
Fixes
Loudspeaker panning gives the same gains on every processor: an arrival on an edge between speakers could leave a speaker beyond that edge a gain of about 1e-17 from rounding (the rear height speakers of 5.x.4, for sound from straight behind, on Intel Macs). Such gains are now 0.
3.2.0¶
On PyPI
pip install acoustic-spacesinstalls the package from PyPI, with the extras in brackets (pip install "acoustic-spaces[viz,gui]"). The GitHub release pages keep the same wheels. The description on PyPI is the README, with its links to the documentation and its picture pointing at GitHub.The repository is now github.com/naelolaiz/acoustic-spaces, the name of the package and of its command; the old address redirects. The documentation, the package’s links and the editor’s About window point at the new one.
A message about a missing optional part says to
pip install "acoustic-spaces[gui]"(with the extra it needs), where it listed the extra’s packages for pip.
3.1.0¶
Tuning the devices for a computer (acoustic-spaces tune, acoustic_spaces.engine.tuning)
acoustic-spaces tune [scene](the concert_hall example by default) times the late stage on the CPU, the GPU and both, for responses of 0.5, 1, 2 and 4 s (--lengths) at the scene’s ray count (--rays), and the wave solver on the CPU and the GPU. It prints the times and keeps the fastest for each in the user cache folder (acoustic-spaces/devices.json);acoustic-spaces infoshows them. In the editor: Simulation > Tune CPU and GPU…, whose window saves on OK only (Cancel saves nothing).A per-computer default device:
simulation.late.deviceandsimulation.wave.devicemay be left out (the editor’s Run on: default), and then follow this computer’s default,cpuunlesstune --default autoor the editor’s tune window (a ticked check box) made itauto;tune --default cpuswitches back. Scenes that name a device keep it. Scenes the editor saved before namecpu; the tune window offers to let the open one follow the default.The late stage’s
autotakes the choice for the run’s response length. Without one, its trial traces the run’s own length (it traced at most 0.25 s), and its choice is kept for that range of lengths. The 3.0.0 filelate-device.jsonis no longer read.The wave solver’s
autocompares the timed costs for the run’s step count, where it took any hardware GPU; withouttuneit still does.
Warnings, errors and finished runs
The editor shows warnings in amber and errors in red, in Messages, in the scene’s form and (the number of problems) in the status bar. On the command line,
error:andwarning:are red and yellow in a terminal (NO_COLORturns that off).When a run, a coverage map or a tuning ends while another window is in front, the editor’s taskbar entry flashes (the Dock icon bounces on macOS).
Starting the editor
acoustic-spaceswith no command opens the editor (acoustic-spaces edit), with theguiextra and a display; without them it prints the short usage and why, as before.-hstill lists the commands.
Tips in the editor
Resting the mouse on a label, field, button or column header of the editor shows what it means, with its unit: the Results table’s rows say what each parameter measures, its band columns which frequencies each octave band holds, and Band which plots it chooses for. Menu entries and toolbar buttons say what they do (with their shortcut), and the scene tree’s entries what they hold.
Fixes
A binaural output through a SOFA file (
{format: binaural, hrtf: ...}, orsimulate --hrtf) gave no IACC_E and IACC_L inparameters.json, the printed table or the editor’s Export; only the spherical head gave them. They now come from the first binaural output, whichever head it has. Such a response is written as.binaural-<file name>.wav, andauralize --format binauralfinds it when it is the only binaural one, instead of asking for that name.Panning onto 5.x.4 and 7.x.4 layouts was lopsided. Their four height speakers, and the rear corners of 7.x.4, each lie on a plane, which was split along one diagonal, so 20 to 28 % of all directions panned differently from their mirror image: a sound from straight above played from two diagonally opposite height speakers, and the reverberation of a symmetric room reached mirror-image speakers up to 2.8 dB apart. Four or more speakers on one plane now get an imaginary speaker at their middle that they play alike.
Responses rendered at 44.1, 22.05 or 88.2 kHz, where 1 ms is not a whole number of samples, had random silent 1 ms gaps in their late part while it was still sparse (in a hall through its first 200 to 400 ms, losing up to about 15 % of the late energy): the noise’s impulses were spread over slightly different slots from those its energy was set in. Rates where 1 ms is a whole number of samples, such as 48 and 96 kHz, were not affected and render as before.
calibrate()tooklf,lfcandt30_x,t30_yandt30_z, which its own error message offered, though the fit cannot work them out: it left them out, or with them alone returned the absorption it started from. It refuses them now, asacoustic-spaces calibratedid.Material libraries of CATT-Acoustic that give the 8 and 16 kHz values after a colon (
<a125 ... a4k : a8k a16k>), as CATT writes them, could not be used at all, andABS1materials (in fractions from 0 to 1) were left out. Both are read now.acoustic-spaces walkthroughran on every core whatever the scene’ssimulation.performance.threadssaid, and took a negative--threads. Likesimulate, it now takes the scene’s setting unless--threadsis given, and so doessimulate_walk.The editor’s Edit > Add material from a library… failed before showing its file dialog. It opens in the scene’s folder now.
A source with a
delay_msgave no stage support (ST early and ST late) at listeners within 1.5 m of it when the delay was more than about 1.5 ms at 1 m (4.4 ms anywhere): the delay counted as distance. Only the distance counts now.simulate()with some of a scene’s sources or listeners (sources=,listeners=, and socalibrate) drew other random numbers for them than a run of the whole scene with the same seed, and others again when the names came in another order. Seeds go by the scene’s order now, as in coverage maps; runs of the whole scene are unchanged.A negative
simulation.seedor--seedpassed every check, then stopped the run witherror: expected non-negative integer. It is refused at once, naming the seed.view --pathsandexport --pathsdrew and counted every path up to the highest order withinir_length_s, where a simulation withtransition_msset, or with the statistical late method, keeps fewer as early paths (13 of the shoebox’s 63 with the statistical method). They show those the simulation keeps now.Particles that reached their 200th reflection (in
animateandexport --particleslonger than about 1 to 1.8 s) flew on through the walls. They stop there now.On a GPU, a late stage of more than about 1 to 3 million rays per band in a small room overflowed its sums, and was then traced all over again on the CPU. The GPU traces it now.
When the GPU declined a run it had been given (its buffers too small for the results), the CPU traced it while the progress message said the GPU had. The message says the CPU now, and why;
autotakes the CPU for such a run, where its trial stopped the run before. Withboth, a second CPU tracer took the GPU’s parts beside the first, on the same threads; the one CPU tracer takes them over now.tunerefuses such a scene, naming the reason, where it timed the CPU as the GPU, andengine.ray_device.tracereturns the plan it followed with the results.The leak finder’s area for a gap whose open edges do not close up into loops, or for rays slipping between faces, came out too small wherever rays escaped elsewhere too (by 23 % for two gaps of 30 m2 in a model of 100 m2). It is the gap’s share of the reflections times the area of the surfaces now.
engine.wave.wave_responsestookmax_band_hzup to 8000 Hz, as its error message said, and failed with anIndexErrorat 8000. Like the scene, it takes 63, 125, 250 or 500 Hz now (WAVE_BANDS_HZ).compare_parameterscompares IACC_E and IACC_L when both sets have them, with the just noticeable difference of 0.075 that ISO 3382-1 gives in annex B (the module’s comment said the standard gave none).A Collada file with
<up_axis>X_UP</up_axis>was read as if it were +z up, lying on its side, andup_axiscould not correct it. It is turned upright now, and an<up_axis>that is none of X_UP, Y_UP and Z_UP is refused, where it was read as +z up.calibrate(bounds=(0.0, high))stopped with an error from scipy about infs or NaNs whenever a fitted material absorbed 0.01 or less in a band, as the built-in brickwork and concrete do at low frequencies. It refuses a lower bound of 0 at once, saying that the bounds must be 0 < low < high < 1.acoustic-spaces inspectof a model whose walls are drawn as a closed shell inside another gave the volume of the cavity between them, while the reverberation times below it used the air inside. It also gives the “air volume” now.A walk with places less than 10 ms apart (a
walkthrough --stepof about 1 cm at walking pace, or a walker faster than 25 m/s) stopped with an error about late parts, or with--late-step 0played some places through their neighbour’s response: the places were named by their time to 10 ms, and shared names. They are numbered now (<name> #<place> @ <time> s). A walk whose points are all where it starts is refused, naming the walker, where it divided by zero.With the statistical late method every place of a walk has a late part of its own, but
WalkResponses.late_placesand thewalkthroughsummary said that only those every--late-stephad. They say every place now.walkthrough --hrtfdid nothing unless the format was binaural, and took a missing file without a word. Likesimulate --hrtf, it now renders binaural through the HRTF when--formatis not given, and refuses another--format.A coverage map over a slab whose top and underside are in one group (an IFC slab read without
space:, or a solid floor from a CAD program) could put its pointsheightabove the underside, up to the slab’s thickness too low. They stand on its top now.The STI of a coverage map point, and in
parameters.json, was 0, or too low, where the decay of a band from 125 Hz up had not reached -20 dB by the end of the response (a shortir_length_sormap --length), while the other parameters there were undefined. It is undefined too now (-, or null in JSON).validate,simulate,mapand the other commands that read a scene printed a Python traceback for a scene file that is not valid YAML, for a folder, and for an IFC or SketchUp model without its extra. They print anerror:now, and the editor says why it could not open a file that is not valid YAML.A scene saved with an inline array whose element is a file (
array: {element: {file: cabinet.csv}, ...}) kept that file’s absolute path, so its folder could not move, and the editor’s Save As with copies left the file behind. It is written relative to the scene’s folder and copied along now.The editor’s floor plan table wrote every corner back as it showed it, to six significant digits, when one changed or a corner was added or removed; the simulation form wrote back every setting when one changed, rounding the response length to 10 ms and the early-to-late time to whole milliseconds. They write only what changed now.
A scene’s
simulation.early.coherent_max_band_hz: null(early paths in energy throughout) was left out when the scene was saved, and when the editor changed any other setting, so it came back as the default, 125 Hz. It is kept now.Scene.to_data()gives a scene as the data its file holds.The coverage form’s note said the sources’ sound power is spread evenly over the bands. It says how the map spreads it: as each source’s spectrum, gain and equalizer set.
The warning for edges that join faces of opposite orientation said that it only mattered for the volume. The volume and the inside checks do not depend on the normals; only their reported direction is unknown, and the warning says so.
The documentation of
Calibration.uncertaintyand thecalibratetable’s+-column now says what it is, the standard error of the absorption for misses of one JND (how far it can move before the sum of the squared misses grows by one JND squared), and the module summary ofrendersays the real-time convolver’s partitions grow along the response, from the block size up to 8192 frames.The gallery example’s comment said that its inner corner is where wall_4 and wall_5 meet; it is wall_3 and wall_4.
The user guide named the editor’s menu entry Edit > Add material (it is Add material from the database) and said that arrays do not walk (only lines and surfaces do not; an array source walks like any point source), and gave the seminar room’s 8 kHz reverberation time as 17 % short (18 %) in the octave bands.
The validation report quoted the statistical late method’s G as within 0.5 dB of the stochastic one (the test allows 0.7 dB), the BRAS octave analysis as within 4 % and 0.5 dB of the published values (CR3’s EDT at 500 Hz is 4.3 % lower, its C80 0.6 dB higher), the seminar room’s 8 kHz shortfall as 17 % (18 %), the wave solver as leaving partitions out (it stops with an error) and the whole-room check as run with default settings (the wave solver is switched on). Its GPU timing table now says it comes from an earlier version of the script, and a paragraph that rendered as table rows is a paragraph again.
CI runs the tests with
pytest -W error, as the contributing guide said it did, so a warning fails the run; ruff runs on Linux.
3.0.0¶
Validation against two measured rooms, third-octave bands, vector-based scattering, sources with their own spectra and line, surface and array sources, sound through walls and between coupled rooms, fitting materials to measurements, ray tracing 7 to 8 times faster, the late stage and the wave solver on a GPU, edge diffraction over two edges, early reflections in pressure at low frequencies, IFC and Collada models, material libraries, a leak finder and walkthroughs.
Simulations scatter by vector and by the size of each surface, and trace rays of their own for each band, so rooms with smooth walls decay faster than in 2.0.0 and closer to measurements (see Scattering for version 2’s settings). Early reflections add in pressure at 63 and 125 Hz, and the wave solver’s walls absorb as their materials do in every band, which changes results in those bands. Scene files of 2.0.0 load unchanged.
Comparing with measurements (acoustic-spaces compare, acoustic_spaces.compare)
A simulated and a measured impulse response, each at its own sample rate, compared by their ISO 3382-1 parameters band by band: both values, the difference, and the difference in just noticeable differences (ISO 3382-1:2009, table A.1), marked where it is more than one; per parameter, the bands within one JND and the mean absolute difference from 125 Hz to 4 kHz.
Bands that cannot be compared (a decay of less than 20 dB, a decay time the noise leaves out, G without a reference on both sides) are listed with the reason, not counted as zero.
A channel and a reference for G per file,
--json, and--plotwith the two decays in each octave band (plot_comparisoninacoustic_spaces.viz).
More parameters
LF, LFC and LJ (ISO 3382-1, annex A) from the directions the energy arrives from, with a figure of eight whose null points at the source; IACC early and late (annex B) from the binaural response when one is written; stage support ST early and ST late (annex C) for listeners within 1.5 m of a source; and the T30 of figures of eight facing forward, left and up from the listener.
STIPA next to the STI (IEC 60268-16), U50, and the echo criteria of Dietsch and Kraak for speech and music.
parameters.jsonhas them for every response, at the source’s level and the coverage noise, andsimulateprints them.analyze --lateral CH,--ambisonic(first-order AmbiX) and--binauralgive the same from measured responses, and--stagethe stage support;analyzeprints the STI, STIPA and echo criteria of each channel.comparetakes LF, LFC and the directional decay times where both responses have them.The editor’s results table shows the further parameters where a response has them.
Validation against measured rooms
Two rooms of the BRAS benchmark, a seminar room and a chamber music hall, as scenes in
validation/bras, with their surface data and the parameters of their measured responses (CC BY-SA 4.0, like BRAS).validation/bras_validation.pycompares simulations with them; the validation report has the results. It prints how far each room has got on standard error.The cross-check against pyroomacoustics averages five runs of pyroomacoustics per room (
--pra-runs), which draws its rays afresh on every run, and shows the spread between them. One run had made its C80 in the lecture hall look up to 3.5 dB lower than this package’s; the mean of five is 0.3 to 1.7 dB lower.
Scattering (changes results: see below for version 2’s)
Vector-based scattering, as in ODEON: every reflection sends a ray in the direction of the sum of its specular direction and a random Lambert direction, weighted by the surface’s scattering coefficient, instead of specularly or, with the coefficient as probability, diffusely. The diffuse rain sends the scattered share of each reflection to the listeners and the detector spheres count only the rest, so nothing is counted twice. Where the bands’ coefficients differ, each band is traced with rays of its own, so a wall can scatter the bass and reflect the treble. Particles in
animateand the editor move the same way.Scattering from the size of a surface (Christensen and Rindel, 2005, as in ODEON): a surface that is small for the wavelength, far from where the sound came from or met at a grazing angle scatters more than its coefficient says. Each flat surface (the faces of one plane joined at their edges, of any material) counts as the rectangle with its area and perimeter. Early reflections lose that share too, and the late part takes it up. Reflections that edge diffraction already completes keep their own coefficient: the first two orders with
first_orderdiffraction (the default), the first withdirect.simulation.scatteringpicks the model:method: vector(default) orrandom,from_size: true(default) orfalse, andby_band(below); the editor’s simulation page has all three.simulation.late.rays(andcoverage.rays) now count the rays per octave band, 50,000 by default instead of 100,000 shared by all bands. The church, concert hall and orchestra examples trace 25,000 per band. A simulation takes about 1.5 to 4 times as long as before;simulation.scattering.by_band: falseturns one ray for all bands by their mean scattering instead, which is faster but scatters the treble like the bass.Rooms with smooth walls decay faster than with version 2: in the measured chamber music hall of BRAS the reverberation time came out 20 to 32 % long and now 7 to 20 %. The shoebox example’s T30 at 1 kHz drops from 2.4 s to 1.3 s. Scene files need no change; to get version 2’s results, set
simulation: {scattering: {method: random, from_size: false}, late: {rays: 100000}}.
Third-octave bands
simulation.bands: third_octaveruns a simulation in the 24 third-octave bands from 50 Hz to 10 kHz instead of the eight octave bands: materials, air absorption, source directivity, the scattering of surface size, the rays (raysper band), the wave solver’s bands and the impulse response’s filter bank all follow them. It traces three times the rays and takes about three times as long.simulate --bandsand the editor’s simulation page set it.Materials take 8 or 24 absorption and scattering values. An octave value holds for its three third octaves; an octave run takes the mean of a material’s three.
Coefficient files with more than one value in some octave (third-octave tables, as in BRAS) keep their third octaves instead of being averaged onto octaves, so such a material’s
absorptionnow has 24 values;Material.coefficients(bands)gives them in either bands. Octave runs average them as before. A band with no data now copies the nearest third octave rather than the nearest octave: a table that starts at 100 Hz gives the 63 Hz octave its 100 Hz value instead of the 125 Hz octave’s mean.Room parameters, the STI, coverage maps and plots stay in octave bands, each octave the mean of its three thirds’ energies.
Response.energy()has the 24 bands andResponse.octave_energy()the octaves;Response.band_hznames them.The editor shows a material with third-octave data in 24 columns, and
inspectits octave averages.In the BRAS seminar room, third-octave bands bring the reverberation time at 250 Hz and 8 kHz from 13 and 17 % short to within 4 % of the measurement; the validation report has both rooms (
bras_validation.py --bands third_octave).
Sources
level_dbtakes the sound power per band (8 or 24 values) as well as in all, andspectrumspreads a level in all evenly (flat, as before) or as male or female speech (IEC 60268-16). The STI takes the female weighting forspeech_femalesources, in responses and coverage maps;speech_transmission_indexhas atalker.gain_db,eq_db(per band) anddelay_msfeed a source: the gain and equalizer act on its sound power and its impulse responses, the delay on every arrival, and the coverage map of all sources together takes each source’s delay. G does not change with the feed (Response.fedapplies one to a response).Sound pressure levels:
simulateprints them and writesspl_dbandspl_a_dbfor each response,Response.levelsgives them for any sound power, the editor’s results have an SPL row, andinspectlists each source’s sound power per band. The editor’s source page sets the spectrum, gain, equalizer and delay.Line and surface sources:
line_morsurface_m(andspacing_m) make a source of uncorrelated point sources that share its power and rays, in responses, coverage maps, early paths and particles. Alambertdirectivity radiates as a vibrating surface. The editor sets the shape and size, and draws them in the 3D view and the plan; the Blender export andinspectshow them;directivity lambertprints the pattern’s index.Coherent arrays:
directivity: {array: ...}, or a YAML file with the same fields, gives a source the far-field pattern of elements fed alike, with their positions, aims, gains, delays and polarities, a column’s count and spacing, and electronic steering (steer_deg). The energy is the exact mean over each third-octave band, with the room’s speed of sound (array_directivity,read_array). The editor andacoustic-spaces directivitytake array files.The pattern is the array’s far field, from about its length squared over the wavelength. The room warns about every listener nearer than that at 1 kHz, naming the array and the distance (
array_length,far_field_distance).
Walls, doors and openings
Materials take
transmission(the share that passes, 1 for an opening) ortransmission_loss(the sound reduction index R in dB), one value or one per band. What passes counts as absorbed on its side, so the absorption must be at least the transmission; without absorption values it is the transmission.A model can hold several rooms, found from the faces themselves: every space they close off, whichever way the normals point. Faces between two rooms, and panels standing in a room, reflect and absorb on both sides. The inside of a thick wall or a column is solid unless a source or listener stands in it.
inspect,validateand the editor list the rooms;Room.room_volumes,Room.room_ofandRoom.face_roomgive them to scripts.Early reflections pass through partitions with the share that passes, and through openings unchanged; late rays go on behind a partition as a copy carrying that share, with a weight window per room from the rooms’ diffuse power balance. Coverage maps cover every room. The statistical late method follows coupled-room theory room by room; the statistical reverberation times count both sides of the faces between rooms.
Openings are holes for diffraction (the edges of a doorway diffract), for the wave solver’s air and for particles; particles pass partitions as often as the sound does. The wave solver refuses walls that let part of the sound through.
A warning for listeners in a room no surface joins to the source’s.
In the editor, a material’s Lets through (nothing, part of the sound with R per band in the table, or everything) and the rooms under the scene’s Properties.
Checked against diffuse-field theory: the level behind 30 and 50 dB partitions within 0.3 dB, an opening the size of the wall the same as one room, and the double slope of coupled rooms within 1 dB of the two-room model; the statistical method and coverage maps across rooms against the same model, the rays and listeners.
Audience seating
audience: trueon a material: sound that reaches a listener low over it loses more in the 500 Hz to 2 kHz bands, after the 2025 fit of Jurkiewicz et al. to scale-model measurements (acoustic_spaces.engine.grazing). It applies to the last leg of the early paths to listeners and map points over the seats. The editor’s Audience seating sets it, andinspectsays so.
Fitting materials to measurements (acoustic-spaces calibrate, acoustic_spaces.calibration)
The absorption of a scene’s materials fitted, octave band by octave band, to the T30, EDT and C80 (or other parameters) measured at its source and listener pairs, in JNDs: from impulse response WAVs,
analyze --jsonreports or a JSON of parameters. It prints the fitted absorption, what it started from and how far the measurements leave each value open, with the errors before and after, and writes the fitted scene.The materials keep their proportions unless the measurements tell them apart: a material moving by a factor of 2 more than the others costs one JND.
Each round traces the room once, and the fit re-weights the rays traced for other absorption (
acoustic_spaces.engine.reweight): the tracer counts each ray’s reflections off each of up to 16 groups of faces (simulate(..., groups=...),Response.late_moments,Response.reweighted).In the editor, changing a material’s absorption after a run updates the results at once by re-weighting that run’s rays (
reweight.reweighted_result); the results panel says so, and Run traces afresh.Checked on the simulator’s own responses: absorption is recovered within 0.05, and a shoebox’s walls within 0.015 from 500 Hz up. Also checked on the two BRAS rooms (
bras_validation.py --calibrate), fitted to one loudspeaker position from BRAS’s initial absorption estimates. On the other position’s responses, both rooms come to about one JND. That is as close as BRAS’s own fitted absorption gets in the seminar room, and closer in the hall: T30 within 0.5 JNDs against 1.8.
Wave solver: walls that follow their materials, and open models (changes results in the wave bands)
Each material becomes a locally reacting surface whose admittance changes with frequency: a constant and resonant mass-spring-damper branches, fitted so that its random-incidence absorption, counted through the analysis’s band filters, matches the material’s in every wave band (
acoustic_spaces.engine.impedance). One run covers all the wave bands, where version 2 ran one for each set of coefficients, and the absorption changes smoothly from band to band instead of stepping. The largest miss of the fit is in the details aswave_absorption_miss;wave_runsis gone.Round an open model the grid grows by a perfectly matched layer of 8 cells, which takes in the sound leaving it: what comes back is 57 to 74 dB below the sound, where the grid’s edge alone sent back 7 to 14 dB below it.
WaveResponse.signalsholds the one run (samples, 4), andWaveResponse.bandsis a field;band_runsis gone._native.WaveSolvertakes each face’s material, the materials’ constant admittances and their branches, and the matched layer.Checked in
validation/wave_solver.py: modes of a box with frequency-dependent walls (frequencies within 0.25 % of perturbation theory, decay rates within 2 %), the normal-incidence absorption of a duct’s end (within 0.004 of the grid’s theory), the matched layer against a grid too large for anything to come back, and the run count.
Wave solver on a GPU (simulation.wave.device, acoustic_spaces.engine.gpu)
device: gpusteps the wave solver’s grid on a GPU through WebGPU (wgpu: Vulkan, Metal or Direct3D 12, on any vendor’s GPU), with the newgpuextra;autotakes a hardware GPU when there is one. The C++ core still lays the grid out, and one WGSL compute shader steps it, one dispatch per time step; the runs agree with the CPU’s to within about 1e-5 of the largest pressure.cpustays the default.simulate --wave-device, a choice on the editor’s Simulation page, the GPU’s name inacoustic-spaces infoand in the progress line, andwave_gpuandwave_stepsin the details.validation/wave_gpu.pytimes a GPU against the CPU on the example rooms. CI runs the comparisons on Mesa’s software Vulkan driver on Linux.
Wave solver: less dispersion, and faster (simulation.wave.scheme; changes results
in the wave bands of closed rooms)
A second stencil, the interpolated isotropic one: 19 points (the six axial neighbours and the twelve edge neighbours) at its Courant limit, √(3/4). Waves on it are 0.67 % slow at 8 cells per wavelength, within 0.03 % the same in every direction, where the standard 7-point stencil is up to 1.74 % slow along the grid’s axes.
auto, the default, takes it in closed rooms and the standard one round open models, since only that one has the matched layer;isotropicwith an open model is an error.Edge links that reach behind a wall are folded onto the axial link to the far end’s image in the wall, so a plane wall reflects as it does with the 7-point stencil and the cells along it are as stiff across it as along it. A cut link’s losses go to the wall it crosses.
The CPU update streams through runs of interior cells, which the compiler vectorises: 1.7 times as fast with the 7-point stencil, 2.4 times with the 19-point one, with the same results to the bit. The grid’s set-up casts rays along the links only from cells near a face, about 3 times as fast. The concert hall’s wave bands take about as long with the 19-point stencil as they did with the 7-point one.
_native.WaveSolvertakesstencil(7 or 19) and reports it, the details havewave_stencil, and the GPU kernel steps both stencils.Checked in
validation/wave_solver.py: the speed of sound along an axis and along a diagonal from the modes of a rigid box, within 2e-5 of each stencil’s dispersion relation, and every other check with both stencils.
Edge diffraction over two edges (simulation.early.diffraction: second_order)
second_orderadds tofirst_orderthe direct sound over two edges in turn: over a thick wall or barrier, whose two corners share its top, round both corners of a doorway’s jamb, or over two screens in a row, wherefirst_orderfinds nothing or only the way round one corner. Each source and listener get the shortest path over each pair of edges that see each other (acoustic-spaces view --pathsdraws it with its two corners).The wave of the first edge reaches the second as the UTD has it, with the second edge’s distance parameter from the curvature of that wave where it gets there: near the first edge’s boundaries it is still curved like the source’s, which the textbook formula for parallel edges misses by up to 6 dB at low frequencies. Along a face the two edges share, the incident wave and its reflection off the face are one wave, and the second edge’s coefficient is halved.
Where the first edge’s wave is lit at the second edge (or the second’s at the listener), the path continues a path over one edge and takes its cross term with that path, as a path over one edge does with the specular path it borders: across the plane of a barrier’s top the level changes by at most 0.8 dB at 63 Hz and 0.1 dB from 1 kHz up, where
first_orderdrops from the path over one corner to nothing.Checked against the exact Biot-Tolstoy-Medwin solution in
validation/edge_diffraction.py: over barriers 0.3, 1 and 3 m thick within 0.35 dB from 500 Hz up and 2 dB at 63 Hz, and over two thin screens in a row within 0.75 dB from 500 Hz up and 2.7 dB at 63 Hz. The validation report has the tables.The early stage of a hall with 16 edges that diffract takes 170 ms per source and listener with
second_order, 120 ms withfirst_orderand 90 ms without diffraction.first_orderstays the default.
Early reflections in pressure at low frequencies (simulation.early.coherent_max_band_hz;
changes results at 63 and 125 Hz)
Up to the 125 Hz band by default (the 63 and 125 Hz octaves, or the six third octaves up to 160 Hz), the early paths add in pressure, not in energy. Each carries the phase of every reflection on its way, that of a locally reacting surface with the material’s absorption (R = (cos θ - Y) / (cos θ + Y), with the admittance fitted as for the wave solver), and a path over an edge the whole complex UTD coefficient, whose terms then add to the specular paths as the theory has them. Above those bands each path keeps the phase it has at their top, so the bands join without a step.
The echogram reads those bands’ early energy from the paths’ pressure, through the band filters of the analysis of measured responses; coverage maps and the material fit do the same. The parameters of a simulated impulse response and of its echogram now agree there: in a 7.3 x 5.1 x 3 m box, G within 0.01 dB and C50 within 0.05 dB, where adding the paths in energy left them up to 2.1 and 1.6 dB apart.
Next to a hard floor the direct sound’s pressure doubles at low frequencies (+6 dB where energy gave +3 dB), and an absorbing floor at grazing incidence takes from it; a loudspeaker a metre out from a wall loses bass at 63 Hz.
Over edges, against the exact solution (
validation/edge_diffraction.py), the levels at 63 and 125 Hz come within 0.05 dB over a thin screen (0.9 dB in energy) and within 0.6 dB over two edges (2.7 dB), and step by at most 0.01 dB across a barrier’s top (0.8 dB).The late part still adds in energy, and so do the uncorrelated points of a line or surface source. With the wave solver up to that band or higher, its bands take their place.
nulladds the paths in energy in every band, as before.--coherent-bandsets it for one run ofsimulateormap, and the editor’s simulation page has it (“In pressure up to”).EarlyReflections.spectrumholds each path’s pressure at the frequencies the edge diffraction is evaluated at, relative to its level and delay;acoustic_spaces.engine.coherenthas the rest.
Leak finder (validate --leaks, view --leaks, acoustic_spaces.engine.leaks)
Rays traced from the sources with diffuse reflections show where a model that is not closed lets them escape. Each gap is reported with the rays that escaped through it, its area and size from the open edges round it, where the rays crossed it and the surfaces next to it. Openings to the outside (surfaces that let sound through) are told apart from leaks, and a source outside a closed model is named.
view --leaksand the editor’s View > Show leaks draw the open edges, the escaping rays and a marker on each gap (plot_leaksinacoustic_spaces.viz); the warning about a model that is not closed says how to find its holes.
Material libraries of CATT-Acoustic and ODEON
A material takes its coefficients from an entry of a CATT-Acoustic file (
ABSlines, in a geometry or library file, with scattering) or an ODEON library (Material.Li8):filenames the library andentrythe material, by name or by ODEON’s number.acoustic-spaces materials --filelists a library’s materials, and the editor’s Edit > Add material from a library adds one.ODEON’s own
Material.Li8is read as ODEON writes it (checked against the library from odeon.dk, 198 materials): the product name of{Product name*=...}names a material and{Category=...}groups the listing, and materials from ODEON’s Material calculator, with its settings and third-octave values on their lines, are read by their octave values.
IFC and Collada models
IFC files (BIM models from Revit, ArchiCAD and others) with the
ifcextra (IfcOpenShell, LGPL): every building element with its openings cut, or, withgeometry.space, one room from its space boundaries, a closed room with a face per surface (from the space’s solid where the file has no boundaries). Faces are named by their element’s name, type, materials and IFC class, and the file’s units are used. The editor’s geometry panel picks the room.Collada (
.dae, SketchUp’s export among others), read directly: triangles, polylists and polygons, nodes with their transformations and instanced nodes (SketchUp’s components), labelled by material and geometry name, in the unit and with the up axis the file records.DXF polygon meshes and closed 3D polylines are read too.
Assignments and material names match surface names whatever their case.
acoustic-spaces examplesand the editor’s examples menu say which extra an example needs when it is not installed.
Walkthroughs (acoustic-spaces walkthrough, acoustic_spaces.walkthrough)
A listener or a point source walks through the room along the corners of its
walkin the scene, at a walking pace, facing the way it goes and turning through corners at 90 degrees per second (turn_deg_s). Responses are simulated every 25 cm along the way; the late part, which changes slowly, is traced every 2 m and mixed in between, so the gallery’s walk of 101 places costs six times one response.Dry signals play through the response of the place the walker is nearest to, with 50 ms crossfades from each place to the next, in any of the output formats.
--videosaves the room and the walker going along its path (with an arrow and its yaw for the way it faces) as a GIF or MP4, in time with the sound (animate_walkinacoustic_spaces.viz); an MP4 carries the sound itself when it has one or two channels (mono, stereo, binaural).The viz extra installs imageio and imageio-ffmpeg, which MP4 videos need (for
animatetoo); before, they had to be installed by hand.simulate(..., late_from=...)gives listeners or sources the late part of others, mixed, which is how walks trace fewer late parts.
Examples
gallery_walk: a visitor walks away from the guide in the L-shaped gallery and round the corner while the guide talks.bim_office: an office from an IFC model in millimetres, with a window and a door, read as one room from its space boundaries;make_model.pywrites the model.rehearsal_room: a rehearsal room under a pitched roof from a Collada file in inches, as SketchUp exports it;make_model.pywrites the model.material_libraries: a seminar room with materials from a CATT-Acoustic file and an ODEON library.leaky_classroom: a classroom model with a missing window pane and a ceiling 3 cm short of a wall, for the leak finder.listening_room: a living room with a loudspeaker against the front wall or a metre out from it.bass.pyprints G at the sofa in the lowest bands from both, with the early paths in pressure and in energy: 5.7 dB apart at 63 Hz in the early part in pressure, within 0.7 dB in energy.archway: a bar and a dining room joined by an archway in a wall 40 cm thick, withsecond_orderdiffraction: a diner behind the wall hears the singer over both corners of the arch’s jamb.town_hall_pa: a sound system for speech in a reverberant town hall, two columns (coherent arrays) beside the stage and two delay loudspeakers with their own gain, equalizer and delay, over seats marked as audience.delays.pymaps the STI with and without the delays: the back rows’ median goes from 0.52 to 0.59, the front’s and the middle’s down by 0.04.workshop: noise from a conveyor (a line source), a machine’s casing (a surface source) and a compressor, each with its own sound power spectrum, mapped as the A-weighted level over the floor.apartment: a living room and a bedroom with a wall and a door between them; the television on the sofa at 73 dB(A) is 45 dB(A) in bed with the door closed, 63 dB(A) with it open.room_modes: the bass of a small bare room from the wave solver, in third-octave bands, on the CPU or a GPU, against rays alone.fitting: a seminar room that measured longer than designed, as the absorbing panels of its back wall were left out.make_measurements.pymakes up the measurements by simulating the room as built;compareshows the gap, andcalibrate --materials panelsfinds the plain wall from 500 Hz up.orchestra: an anechoic recording of Beethoven’s 8th symphony (TU Berlin, CC BY 4.0) on the stage of the concert hall, one source for each of its 61 tracks, with scripts that mix a movement at a seat, draw pictures of the hall and the seating, and make a video of the mix. The README links the video made with them.Scripts in the example folders (
.py,.sh) are installed with the scenes.
SketchUp files (experimental)
.skpmodels read directly with thesketchupextra, through openskp (MIT), which needs no SketchUp: faces labelled by their material, matched by colour. The editor opens them too. OBJ exported from SketchUp stays the recommended path.
Smoother binaural
HRTFs are interpolated between the measured directions instead of taking the nearest: barycentric weights on a triangulation of the measured directions (as in VBAP), with the responses aligned at their onsets and the mix delayed by the mixed onsets, to a fraction of a sample. Reflections between measured directions get an interaural time difference between theirs, so moving a source or the listener’s head in a scene no longer makes it jump in steps. A set measured round the horizon only works too.
Delays in SOFA files (
Data.Delay) keep their fractions of a sample.scipy is a new dependency (for the triangulation).
Geometry
Where a face’s edge runs past the corners of faces on its other side (a T-junction, which SketchUp stitches with faces of zero area), cleaning splits the edge at those corners, so the model stays closed.
Rooms touching along an edge, with four faces on it, count as closed when the faces pair up. Edges shared by an odd number of faces still leave a model open, and the report and the warning count them.
The volume and the inside checks no longer depend on which way the normals point: a model with flipped faces has its volume, and points near a flipped face count as inside. A model counts as closed when every face has a room on at least one side.
Head tracking
Turning the sound field is exact and about three times faster (0.3 ms instead of 0.8 ms per orientation at seventh order): each order’s rotation follows from the one below by Ivanic and Ruedenberg’s recursion instead of a least-squares fit on a grid of directions.
Faster ray tracing, and performance settings (simulation.performance)
Ray queries are 7 to 8 times faster, and the late stage with them: the model’s bounding volume hierarchy is split by the surface area heuristic instead of at the median, walked nearest box first, and its box test no longer calls the maths library. One band of 50,000 rays in BRAS’s chamber music hall takes 5.7 s instead of 41 s on four cores. Rays meet the same surfaces at the same distances; where a ray meets two faces at exactly the same point (a shared edge), it may now take the other.
simulation.performance.threads(default 0: every core) andray_tree(sah, the default, ormedianfor comparison) in scenes, under Performance on the editor’s Simulation page, and as--threadsand--ray-treeforsimulate,mapandcalibrate(--threadsforviewandexport). Without--threadsthey take the scene’s setting;infoprints the number of CPU cores.The ray searches run on a GPU too, through WebGPU (
gpuextra): the first part of tracing rays there.validation/rays_gpu.pytimes them against the CPU on rays like the late tracer’s and checks that both find the same faces.Bvh.layout()hands the tree to it, andBvh.occludedtests many rays at once on the CPU.validation/wave_gpu.pytimes the grid’s set-up and the time steps apart (the speed-up is that of the steps), starts the GPU before timing, prints its settings and each grid’s cell size and stencil, and lines up its table in a terminal.The late stage’s rays can be traced on a GPU (
simulation.late.device:cpu, the default,gpu,bothorauto;acoustic_spaces.engine.gpu_tracerandray_device). One WGSL compute shader traces what the CPU does, in single precision and with random numbers of its own, so the two agree statistically.bothshares the rays between the CPU and the GPU as each is free;autotimes a short trial on both and keeps the fastest, or the CPU when wgpu finds no hardware GPU or nothing is clearly faster, and saves its choice for the computer. Sound through partitions and calibration’s face groups stay on the CPU.simulate --late-device, “Run on” on the editor’s Simulation page, and--no-device-trial(simulate,edit) to take the saved choice without a trial.validation/late_gpu.pytimes the late stage on each device and compares the parameters it gives.Tests hold the ray tracer to results stored from main before these changes (
tests/reference/tracer.npz): the trees’ faces and distances, specular paths, and every kind of late run on the CPU (to single precision), and statistically on a GPU.
Fixes
The late stage’s count of rays that left an open model (
escaped_raysin the run details) was always 0: a ray whose search found no face within the distance left to it was taken for one that had run out of time. A ray that finds none is now checked for a face anywhere ahead, on the CPU and on the GPU.Runs that split rays (the late stage’s weight window, sound passing both ways through a partition) came out differently when the core was built with Clang, as on macOS, than with GCC or MSVC: the two numbers that seed a copy’s random stream were drawn in an order the compiler chose. They are now drawn in the order GCC and MSVC used, so Linux and Windows give the same results as before and macOS gives those too.
The sound diffracted by the free edge of a thin screen left out the term of the reflection off the screen’s back wherever the reflection off its front reached the listener: up to 0.9 dB too little there, at high frequencies.
The hint for a missing extra names the packages to install. It said
pip install "acoustic-spaces[extra]", which looks for the package on PyPI, where it is not.Arrays, whose patterns are worked out in third octaves, were shown band by band as if those were octaves:
inspect,directivityand the editor listed 24 directivity indices under the 8 octave bands, the polar plots drew the lowest eight third octaves labelled as octaves, and the 3D view’s balloon showed 125 Hz for 1 kHz. They show octaves now. Simulations were not affected.The editor’s Stop button for playback was easy to miss: it sat at the end of the Results panel’s bar, out of sight once another panel was on show, and the toolbar’s Cancel, which only cancels a run, had the same stop icon. While a recording plays, the status bar now says so, with Stop playback, which is also in the toolbar and the Simulation menu; Cancel has an icon of its own, and Esc stops playback when nothing is being placed or drawn. Moving a source or listener while the sound played looked as if it changed what plays, which it does not: the status bar marks the sound as from before the latest changes, and the log says how to hear them.
Picking another source or listener, or ticking Head tracking, while a recording played in the editor left the sound as it was. The playback now goes on from the same point with the response picked (the run has them all), at the same level, so the difference between listeners stays audible (lowered only if it would clip).
The editor’s Listen played mono when the scene wrote no binaural output, while head-tracked listening is always binaural, so the two sounded nothing alike (the reverberation above all), and each set its own level. Listen plays binaurally now, through the spherical head unless the scene names an HRTF, and both play the dry signal at half its level through the response’s omnidirectional energy (Listen lowered only where it would clip), so ticking Head tracking changes only how the head turns the sound.
2.0.0¶
Coverage maps, edge diffraction, drawing rooms in the editor, head tracking and a wave-based solver for low frequencies.
Simulations now include edge diffraction by default, so rooms with convex corners or
thin screens sound and measure differently than in 1.1.0. Set
simulation.early.diffraction: none (or --diffraction none) for purely specular
early reflections as before. The wave solver is off unless switched on.
Coverage maps (acoustic-spaces map, the editor’s Simulation > Run coverage map)
Room parameters on a grid over the audience area: SPL and G, C50, C80, D50, EDT, T20, T30 and Ts per octave band, the A-weighted level and the speech transmission index (IEC 60268-16, with masking, reception threshold and background noise).
Maps for each source and for all sources together, from the sources’ sound power.
Several levels of seating under one another (stalls and balcony); columns along a slanted wall keep the part over the seating.
The late part is one ray tracing run per source with a track-length estimate through the slab of air at each point, so a map costs about as much as a few listeners.
A top view with the values (
plot_map,--plot), coloured tiles in the 3D view (plot_room(coverage=...),--screenshot) and every value as JSON (--json).The scene’s
coverage:settings (surfaces, height, spacing, rays, noise), with a form in the editor; the concert hall, lecture hall and church examples set them.
Edge diffraction (simulation.early.diffraction, --diffraction)
Sound bends round edges with the uniform theory of diffraction (UTD), averaged over each octave band: the direct sound and first-order reflections (
first_order, the default), the direct sound only (direct), or none.Edges are found in the model: convex corners with more than 195° of open air and the free edges of thin screens; collinear pieces of one edge are joined.
In the shadow a diffracted path carries the sound; on the lit side the specular path next to the boundary takes the interference, so levels stay continuous across shadow and reflection boundaries.
Simulations, coverage maps,
view --paths, the Blender export and the editor (a Diffraction choice on the Simulation page) all use it; diffracted paths are drawn in their own colour and counted in the results.A new example,
gallery: an L-shaped room with a listener round the corner.
Drawing rooms in the editor
A Plan tab next to the 3D view: the room from above on a grid, with every corner of a floor plan or box room as a handle. Drag a corner to move it (a box becomes a floor plan with the same walls), double-click a wall to add one, right-click to remove one.
Draw room (and File > New room from a floor plan): click the corners of a new floor plan. Sources and listeners left outside move inside; over a model file the editor asks first.
Sources and listeners move by dragging, in the plan and in the 3D view, keeping their height above the floor; Shift moves them up and down in the 3D view.
Snapping to a grid (Shift for none), wall lengths while dragging, and refused moves when walls would cross.
plan_probleminacoustic_spaces.geometry: floor plans whose walls cross, touch or run back on themselves are refused with the reason.
Head tracking (play --head-tracking, the editor’s Head tracking controls)
Binaural playback of an Ambisonics response that turns with the listener’s head, in real time: the sound field is turned once per block, crossfading from the previous orientation, then decoded to the ears; the latency is still one block.
Head trackers send OSC over UDP:
/ypr,/quaternion(/quaternions,/quat),/yaw,/pitchor/rollunder any prefix;--invertfor trackers that count an axis the other way round, and a centre to take as straight ahead.In the editor, the Results panel’s Head tracking controls turn the head by hand or from a tracker, and Listen plays head-tracked.
binaural_decoderinacoustic_spaces.spatial.decoder: MagLS (or least squares) filters from Ambisonics of any order to the ears, from any HRTF.FieldRotationand rotation helpers inacoustic_spaces.spatial.rotation: matrices that turn Ambisonics sound fields of any order.render.playtakes any block processor, such asHeadTrackedBinaural.
Wave-based solver for low frequencies (simulation.wave, simulate --wave)
The lowest octave bands come from the sound field itself instead of rays: finite differences in time on a cubic grid over the room (the standard leapfrog scheme at its stability limit), in the compiled core, on all cores.
Room modes and sound bending round everything at low frequencies; the response joins the ray-traced one through the synthesis’s crossovers, from 22 Hz up to the top of the highest wave band (
max_band_hz: 63, 125 by default, 250 or 500).Surfaces absorb as their materials do in each band (locally reacting, with the admittance whose random-incidence absorption is the material’s); bands with different coefficients get runs of their own on the same grid.
Closed rooms, open models (the air at the edge of the grid absorbs) and thin screens; a listener outside the source’s air is refused with its name.
First-order directions from the pressure gradient round the listener, so Ambisonics, binaural and loudspeaker outputs keep the low end’s direction.
The editor’s Simulation page has a Wave solver switch and its highest band;
simulate --wave/--no-waveand--wave-bandoverride the scene.Checked against theory: the modes of a box and their decay rates (within 0.5 %), the reflection off a duct end and the attenuation along absorbing duct walls, the free-field level (within 0.35 dB) and a box’s modal response (
validation/wave_solver.py).
Also
speech_transmission_indexandmodulation_transferinacoustic_spaces.parameters.Parameters from energy responses are computed about twice as fast.
A code review before the release fixed 38 bugs in the simulation, the parameters, the wave solver, the spatial outputs, the geometry import, the command line and the editor. Against 1.1.0, G with the statistical late method rises by up to 1.8 dB, and a response that decays less than 20 dB has no parameters instead of wrong ones.
A scene’s geometry.units now defaults to the units a DXF file records, and to metres
for other files, as before.
Fixes: simulation and parameters
The statistical late method keeps Barron’s total energy: its tail starts at the direct sound and takes off the early paths, instead of starting at the transition and losing the scattered share of the early reflections. Coverage maps use the same tail.
A response whose decay has not reached -20 dB has every parameter NaN, as T20 and T30 already had, instead of values read from the estimated tail.
The noise floor is looked for before trailing digital silence (zero padding, the zeros
remove_latencyleaves), which made T30 read up to 70 % long.Mono WAV columns, as
read_wavgives them, work indeconvolveand the parameter analysis; an empty impulse response has NaN parameters.A run with one late ray, and
simulatewith no listener, give a result or a plain error instead of failing deep inside.Coverage maps no longer fail on points whose response ends within 100 ms of the direct sound.
Fixes: diffraction
A specular path interferes with its own wave’s diffraction terms only, which keeps the level continuous across reflection boundaries (it stepped by up to 2.1 dB).
Paths that reflect off the floor or ceiling where an edge ends are kept once, not twice or never.
A listener or map point on the source no longer divides by zero.
Fixes: wave solver
The wave bands start at emission; they arrived one grid step early.
The listener’s velocity sits on the pressure’s samples (it led by half a step).
The rays’ late noise joins the wave bands in energy, which removes a 3 dB dip at the crossover.
The source adds no net air, which had left a static pressure and a floor about 50 dB down on the wave bands’ decay.
A wall thinner than a cell gives each side its own face.
The grid runs at a multiple of a low output rate instead of with coarser cells, and listeners outside the grid’s air are refused before the runs.
Fixes: spatial audio
HRTFs keep their gain when resampled: a 96 kHz SOFA file used at 48 kHz played 6 dB quiet.
SOFA files that turn the head for each measurement, or give the ears in spherical coordinates, are read with the right directions and ears.
5.0.4 and 5.1.4 keep arrivals from behind on the horizon, panned as in 5.0; they went up to the rear height speakers.
A head path given to
HeadTrackedBinaural.runis not inverted by the tracker’sinvert.Ambisonic encoding refuses directions of zero length instead of giving NaN.
Fixes: geometry, scene files and the command line
Faces of a closed model that point the wrong way are turned to agree with their neighbours, so the inside test no longer puts sources outside the room.
DXF files are read in the units their header records when the scene gives none, and 3DFACE triangles no longer get a fourth corner at the origin.
Polygons that repeat a corner are triangulated correctly: an L-shaped floor had 8 m² instead of 5 m².
OBJ files with several groups on one line, lines continued with
\, or coordinates that are not finite numbers are read or refused properly.A model whose faces all have no area is refused with a clear error.
NaN and infinity are refused in scene files and coefficient CSV files.
A listener within 1 cm of a source is reported as a problem instead of giving a NaN response.
Command-line overrides are checked as the scene file is, and
--sample-ratenow reaches the wave solver.analyze,deconvolveandsweepreport unreadable inputs as errors and create their output folders.
Fixes: editor
Choosing a material for one wall assigns it to that wall, not to every surface that shares a name with it.
The 3D view stays lit after it redraws the room, and keeps its camera while the room is edited.
Typing in a position, orientation or box field changes only that value; the others were rounded to two decimals. Humidity shows a tenth of a percent.
A custom loudspeaker list is shown as it is, without a 5.0 layout added on any change.
Results of a run whose scene was closed are dropped, and saving an example elsewhere keeps its results current.
One recording plays at a time.
Box and floor-plan fields show the units they are in, and boxes may be larger than 1000 units.
Save As adds
.yamlto a bare name, asking before it replaces a file, and opening another scene cancels placing a point in the 3D view.
1.1.0¶
The graphical editor, export to Blender and source directivity.
Graphical editor (acoustic-spaces edit, or acoustic-spaces-editor; the gui extra)
A Qt window with the room in 3D, a scene tree, a form for each item and a message log. Every change is checked against the scene schema and can be undone.
Sources and listeners placed by clicking a floor or wall; surfaces given materials, materials added from the database and their coefficients edited per octave band.
New rooms as a box or from a model file, box and floor plan dimensions, model units and up axis, simulation settings and output formats, all from forms.
Simulations in the background with progress and cancel; results with ISO 3382-1 parameters, echograms, decay curves, arrival directions and reflection paths in 3D; export to the files
simulatewrites; listening to a WAV file through a response.The results show the source and listener selected in the scene, and stay on them after another run. The selected point has arrows to its front and to its top, so yaw, pitch and roll all show in 3D.
Export to Blender
acoustic-spaces exportand the editor’s File menu write a scene as glTF 2.0 (.glb) for Blender: the surfaces with their materials (absorption and scattering as custom properties), sources and listeners facing their way, the early reflection paths by order, and sound particles as an animation in slow motion.
Source directivity
Directivity per octave band: a pattern name or a first-order
afor each band, or measured data from a SOFA file (AES69FreeFieldDirectivityTF, or impulse responses) or a CSV table of levels by azimuth and elevation, interpolated over the sphere.It applies to the direct sound, the early paths, the rays of the late part (in the compiled tracer) and the sound particles, in the source’s own frame including roll.
acoustic-spaces directivityprints a directivity’s index and front-to-back ratio per band and draws its polar patterns (--plot);inspectlists each source’s index. The editor picks a pattern or a file, shows the index and the polar patterns, and draws the selected source’s balloon in 3D. The control room example’s monitors narrow from omni in the bass to supercardioid at 8 kHz.acoustic_spaces.directivityhas the patterns, balloons and file readers;Room.directivitiesholds each source’s.
Examples
New example rooms: a concert hall, a control room and a church.
The examples ship inside the package.
acoustic-spaces exampleslists them and copies them (--copy DIR);acoustic_spaces.examplesfinds them from Python.
Other changes
uv syncinstalls the test tools (adevdependency group: pytest, h5py, ruff), souv run pytestworks in a fresh checkout, and rebuilds the compiled core when the C++ sources change. The README shows the uv and pip setups side by side, and messages about a missing extra nameuv sync --extra ...as well as pip.inspectadds the Arau-Puchades reverberation time, which allows for absorption that faces one way (alsoRoom.reverberation_time("arau_puchades")), and a note when the absorption is too uneven for Sabine and Eyring to hold, as in the example shoebox, whose simulated T30 is several times their estimates.Tests check the decay of a box with specular walls against its closed form, with even and with uneven absorption, and the example’s absorption with diffusing walls against Eyring.
Saved scene files refer to model, material and HRTF files under their folder by relative paths, so the folder can move, and leave out unset fields.
acoustic_spaces.results.write_resultswrites a simulation’s responses andparameters.json, assimulatedoes.
1.0.0¶
The first release: everything the v1 plan set out, from a CAD model to a room you can listen to.
Scenes and geometry
Scene files in YAML with a checked schema: geometry, materials, sources (with directivity), listeners (with orientation), air, simulation settings and outputs.
Geometry from OBJ (with material groups), STL, PLY, OFF, glTF/GLB and DXF (with layers), or built from dimensions: a box or a floor plan raised to a height.
Mesh checks and cleanup: closed, consistently oriented, inside or outside, distance of sources and listeners to the walls.
A built-in database of absorption and scattering coefficients per octave band, plus CSV coefficient files in any band layout.
A 3D view of the room with reflection paths, and animations of sound particles saved as GIF or MP4.
Simulation
Early reflections as exact specular paths, found by deterministic ray tracing or by image sources, then traced back from the listener.
A stochastic late tail by ray tracing (diffuse rain, secondary sources or a detector sphere) with weight-window splitting, or a statistical tail from Eyring and Barron.
A compiled core (C++, nanobind) with a BVH; simulations run on every core.
Impulse responses per octave band, ISO 3382-1 parameters (EDT, T20, T30, C50, C80, D50, Ts, G) and ISO 9613-1 air absorption.
Spatial outputs
Binaural (HRTFs from SOFA files or a spherical head model), Ambisonics (AmbiX, orders 0 to 7) and loudspeaker layouts by VBAP, from stereo to 7.1.4 (silent LFE channel).
Listening
Offline convolution and auralisation of dry recordings, with resampling.
A real-time convolver in C++ with non-uniform partitions, and playback or live input through the sound card (
sounddevice).
Measurement
Exponential sine sweeps, playback and recording, deconvolution with a loopback channel, and the same parameters from measured impulse responses (noise compensated).
Validation and release
A validation report: the automated checks against theory, a cross-check against pyroomacoustics in four rooms, and a benchmark of the real-time convolver (
docs/validation.md, scripts invalidation/).Documentation built with Sphinx; wheels for Linux, macOS and Windows on CPython 3.10 to 3.14; tagged releases publish to GitHub releases and, once set up, to PyPI.