Roadmap¶
v1, version 1.0.0:
M0 Foundations: package layout, compiled core, scene schema, CI.
M1 Scene and geometry: CAD import, material database, 3D view.
M2 Mono simulation: ray-traced early reflections, stochastic tail, IR synthesis, ISO 3382 parameters.
M2b IR measurement: exponential sine sweeps, playback and recording, deconvolution, parameters of measured IRs.
M3 Spatial outputs: binaural (SOFA HRTFs or a spherical head), Ambisonics (AmbiX) up to 7th order, VBAP loudspeaker layouts.
M4 Rendering: offline convolution and auralisation in Python, real-time partitioned convolver in C++, playback.
M5 v1 release: validation report, documentation, wheels and release workflow.
v2, version 2.0.0 (version 1.1.0 had the first three):
A graphical editor: scenes, materials, sources and listeners, runs, results and listening.
Export to Blender: rooms, reflection paths and particle animations as glTF.
Source directivity per octave band: patterns, and measured data from SOFA or CSV files.
CLF loudspeaker files, once their specification is available.
Head tracking: binaural playback of Ambisonics that turns with the head, from OSC head trackers or by hand.
A wave-based solver for low frequencies: room modes from finite differences in time, joined to the rays.
Coverage maps: parameters and STI over the audience area.
Edge diffraction (UTD) of the direct sound and first-order reflections.
Editing geometry in the editor: drawing floor plans, dragging corners, sources and listeners.
v3, version 3.0.0:
Comparing a simulation with a measurement in just noticeable differences (
acoustic-spaces compare).Validation against measured rooms: two rooms of the BRAS benchmark.
SketchUp files read directly (experimental, the
sketchupextra).More parameters: LF, LFC, LJ, IACC, ST early and late, U50, STIPA, echo criteria, directional decay.
Vector-based scattering, and scattering that follows surface size and frequency.
Third-octave bands.
Sources with their own spectra, gain, EQ and delay; speech spectra for the STI.
Line and surface sources.
Loudspeaker arrays with phase (their far-field patterns).
Sound through walls and between coupled rooms.
Audience seating: the attenuation of sound that skims the seats.
Fitting absorption to measured parameters (
acoustic-spaces calibrate); in the editor, results follow a change of absorption at once.Wave solver walls that absorb as their materials do in every band, in one run, and a matched layer round open models.
A GPU path for the wave solver through WebGPU (AMD, NVIDIA, Intel and Apple GPUs alike).
Ray tracing 7 to 8 times faster, and the late stage’s rays on a GPU too, alone or sharing them with the CPU.
A wave solver stencil with less than half the dispersion, the same in every direction.
Second-order edge diffraction: the direct sound over two edges, checked against the exact solution.
Early reflections that add in pressure at low frequencies, with the phase of every reflection and of the sound over edges.
A leak finder: where rays escape from a model that is not closed.
Materials from CATT-Acoustic and ODEON material libraries.
IFC (BIM) and Collada import.
Walkthroughs with moving sources and listeners.
Binaural rendering with HRTFs interpolated between the measured directions.
Version 3.1.0:
Timing the CPU and the GPU on each computer and running on the faster one (
acoustic-spaces tune).
Version 3.2.0:
On PyPI:
pip install acoustic-spaces.
Version 3.2.1:
Wheels for Intel Macs.
Version 3.2.2:
Fix: the editor no longer aborts on looking for a GPU (wgpu’s OpenGL backend).
Translation supported. Spanish ready on catalogs and editor.
Version 3.2.3:
Added Catalan translation.
Added Italian translation.
Added German translation.
Added Dutch translation.
v4, version 4.0.0: outdoor acoustics (ground, wind, temperature, turbulence, vegetation, terrain and buildings from a map, weather classes and long-term levels, Doppler, whole cities, on the GPU), reflections that fade at edges, coherent arrays, groups, movement along keyframes with every mover on one clock, demos of any scene and of the examples, HRTF sets to download, the ISO 3382-2 checks and packs to run on another computer, in the sections below.
Edge diffraction:
Reflections that fade in and out at the edges that cut them off: every specular path, of any order, takes the uniform theory of diffraction’s term of its own boundary at each edge it passes, checked against the exact solution past a screen’s end, a building’s corner, a strip and a gap (within 0.2 dB), and walked round the gallery’s corner 1 cm at a time.
Paths that two events stop at once, as a reflection past the end of its face with an edge in its way: followed where one or two edges explain the events, so the street’s garden no longer steps by up to 1.9 dB in one band (0.63 dB(A)) where the second event clears.
The terms of one event at different edges in the bands that add in energy (indoors from 250 Hz): they take their cross terms with each other, as one wave, so behind the
archway’s wall the level steps by up to 1.4 dB where the reflection off the arch’s jamb appears, instead of 4.4 dB.Reflections cut off by a thick end: behind the archway’s wall the level steps by 0.56 dB at 63 Hz where they go from hitting the jamb’s end face to its far face, a face of the end’s other corner. This and the fade’s other known issues are listed in Validation.
Paths over a thin screen’s edge weighted by the part of their Fresnel zone that the edge covers, as BTM’s integral over a finite edge has it, and kept past its free ends: past a screen’s right-angled corner within 0.1 dB of BTM, where they stepped by 2.8 dB; past corners of 45° to 170° and beside a bend of a screen’s top, each term weighted for its own boundary, within 0.17 and 0.04 dB.
The same at the corners of solid wedges, where three edges meet: sharing the fade as at a screen’s corner came out up to 1.25 dB above BTM at 63 and 125 Hz near a block’s corner, so their paths still stop at the ends.
The same for the edge fade’s terms and for paths over two edges, which still stop at the ends of their edges: in sight just over two barriers in a row the level steps by up to 6 dB where the way passes the end of one (see Known issues).
Reflections on paths over two edges (
second_order): behind a thick barrier on a hard ground the engine has one of the four paths over it, 1.7 to 5.7 dB low from 500 Hz up.The path over two edges that the source’s wave passes in a row, with
first_orderanddirecttoo: without it the level just over a flat roof or along a wall in line with both is too low (6.4 dB at 63 Hz half a metre above a 10 m roof), and behind the roof with the source above its plane up to 6 dB too high (see Known issues).Rows of edges in simulations: outdoors, where sound passes over the barriers and buildings of a scene, the level in the bands up to 500 Hz from the solver for whole rows in place of the paths over one edge and two, in still air and in wind (
simulation.early.rows, see Sound over rows of barriers and buildings).The solver for whole rows in coverage maps and along walks, which take the paths over the edges: a run for every point takes seconds, and along a walk the level would step where the solver takes one place and not the next.
The solver for whole rows where it does not hold yet: rigid screens on a rigid ground (up to 1.3 dB low near grazing incidence, 3.4 dB at street level behind three screens, and up to 8.2 dB where the way crosses them at a slant of 75°; part of what lacks is the sound that goes to and fro between them); an end almost straight under a top, within some 4 cm of its first obstacle, or close before it and high above its top, and ends within half a wavelength of each other (refused); an end in the shadow within decimetres of a top (up to 0.5 dB low); an end behind a change of the ground, on a roof that reaches the other end with nothing on it between them, or with nothing between it and the other end, with a way on or a reflection off the ground or the roof steeper than 35°, or closer than a wavelength to its first top or to the other end (refused); a roof’s edge or a wall close behind an end; barriers lower than a wavelength where the edges’ terms stand, which lack the sound that goes on between a barrier’s top and the surface under it (a parapet on a roof’s edge up to 1.3 dB off; a barrier with anything else between the ends, on a ground with an impedance, or with the ends closer to each other than about a quarter of its height, both within about a twentieth of it from the barrier, or between the two; only a barrier alone on a rigid ground or roof is taken as the strip it makes with its image), and steps; a row that the way crosses at a slant on a ground with an impedance, or with edges in different directions, where an end’s way to it, or its image’s in the ground, is steep or an edge in it that a march lights is turned more than 60° (refused: the row solved across its edges and moved to the slant is for a rigid ground and edges in one direction; not tried yet: a ground with an impedance by its impedance at each wavenumber across the row), edges turned more than 75° (refused: not measured beyond), a row that the sound crosses at wavenumbers below the solver’s lowest (refused), and edges at different slants (not measured); a wall given as a thin building (up to 0.4 dB low); the tops’ place in the march’s cells (up to 0.2 dB); roofs and walls that absorb (rigid in it; a roof with an impedance leaves the row to the paths over the edges), the sound that goes to and fro between a row’s walls, and the bands above 500 Hz.
Paths over edges tested from the edge itself, not from a millionth of the model’s size off it: exactly in the plane of several roofs in a row the level is up to 9 dB below that just above it, and within twice that distance below the plane 9 to 15 dB above that further below.
Corners to the least distance of a ray: the search for early paths takes what lies within its tolerance for planes (a ten-millionth of the model’s largest extent, 1 mm in a model 10 km wide) of the line where faces meet as on that line. A source or a listener within that distance of the line where a wall ends on another that goes on (a T) has reflections that do not exist, off both walls at one point or off both sides of the ending wall there, at up to 8 % of the places 7 m away (none at a right angle) and at up to 3 % at four times the distance; at a plain bend one such reflection, off both walls at one point, stays at 2 of 96 places (seen at bends of 65 and 66.5 degrees, half the tolerance from the line). And two different paths with one image of the source that reflect off the same surfaces in another order count as one: between two rooms as deep either side of a partition, one of 155 paths up to the fourth order (0.1 dB in a band).
The same reflections whichever way round: the rays of the source and those of the listener can find different reflections between two faces, such as two screens or a roof’s wall and a screen level with the roof (up to 0.7 dB apart, and 8 dB where the level is far down), and the edge fade follows the reflections off a roof’s wall one way round only past the end of a screen before the roof (up to 1.3 dB apart, 4 dB with
first_order; see Known issues).Barriers with bends of less than a degree: beside each bend a strip up to centimetres wide has no path over the top or two (see Known issues).
A way exactly through a corner of a flat top, in the top’s plane (no sound over the box alone), and reflections of second order and up that pass level with several tops (5 to 6 dB too high in one scene; see Known issues).
A coefficient for two separate edges in each other’s transition zones, with a transition function of both edges’ parameters: near grazing incidence over two screens in a row the product of two coefficients is 1.0 to 2.1 dB low from 2 kHz up.
Near a shadow boundary, in the bands that add in energy (indoors from 250 Hz), the cross term between the term that continues the geometric wave and the edge’s other terms, weighted by how coherent they are over the band: on the boundary of a right-angled corner the level is 1.7 dB low at 250 Hz and 1 dB at 1 kHz, and on that of a wedge open over 225° 2.2 dB at 250 Hz (see Validation).
Absorbing faces near grazing incidence, in one treatment: reflections and paths over edges that take no step where a source or a listener crosses the plane of an absorbing face, are the same both ways round, are right for a face of any size (a small absorbing patch included) and where two materials share the path, for reflections of any order and in the bands that add in energy as in those that add in pressure; checked against full-wave runs (see Validation, the paths over two edges and the reflections).
Parameters:
The checks of ISO 3382-2 on every T20 and T30: the non-linearity ξ of the line it is read from and the curvature of the decay, amber in the editor past the standard’s limits.
Error estimates, step 1: a material’s absorption uncertainty, and the spread it leads to in the room parameters by Monte Carlo (GUM Supplement 1), re-weighting the traced run, in
simulate’s output,parameters.jsonand the editor’s results.Error estimates, step 2: the rays’ random error in the same Monte Carlo (Student’s t from the batches’ standard error), a combined standard uncertainty and its budget.
Error estimates: default uncertainties for the built-in materials, whose source the database does not name (plus or minus 0.1, or twice ISO 12999-2’s reproducibility).
Error estimates: the number of draws chosen as the adaptive procedure of GUM Supplement 1 does (JCGM 101:2008, 7.9), or given.
Error estimates: each result’s one-sided bounds, for a limit on one side, of 95 % or another probability.
Error estimates: a material’s absorption uncertainty on its page in the editor.
Error estimates: the spread of each listener’s levels and speech results (the level per octave band and A-weighted, STI, STIPA, U50, the level relative to free field).
Error estimates: the uncertainty of a coverage map’s values, its insertion loss and long-term levels too, by the law of propagation of uncertainty (the map traced twice more for each material with an absorption uncertainty).
Error estimates, next steps: the scattering’s uncertainty; a listener’s insertion loss and long-term levels; convergence checks; the method’s own error from the validation.
Outdoor acoustics, in steps:
O1 Outdoor scenes: flat ground under an open sky, buildings, barriers and model files on it; responses as long as the distances need; coverage maps over the ground.
O2 Ground impedance (grass, soil, snow) and the spherical-wave ground reflection; mixed ground in zones, weighted by the Fresnel zone; the early paths in pressure in every band outdoors; the wave solver fitted to the impedance.
O3 Wind and temperature gradients: curved rays and shadow zones; a parabolic equation for the sound over the ground below 500 Hz and in the shadows; validated against measurements of a wind turbine up to 845 m away.
O4 Atmospheric turbulence: the coherence of the paths (the ground dip filling in), sound scattered into shadows, and impulse responses, renders and walkthroughs that flicker as the eddies drift past; checked against a parabolic equation through random eddies.
O5 Vegetation: belts of trees, hedges and lone trees as volumes of foliage and stems that absorb and scatter sound, along the early paths and the late rays; checked against a separate particle simulation.
O6 The wave solver in wind and temperature: a speed of sound per cell (the effective one along the horizontal from the source) in the grid, its walls, its source and its matched layer; checked against the parabolic equation at 63 and 125 Hz. The flow itself (the linearised Euler equations) stays out (O18 adds moving air).
O7 Outdoor results: the level relative to free field and the insertion loss of buildings, barriers and vegetation (a second run without them), at listeners and on maps, in the report, on the command line and in the editor; checked against the exact solutions for a half-plane and for a ground.
O20 Sound reflected off the ground both before and after bending over an edge (the fourth of the four paths over a barrier on a hard ground):
simulation.early.diffraction_reflections: both_sides, on the command line and in the editor; the barrier’s insertion loss is then within 0.15 dB of the exact solution.one_sidestays the default, so existing scenes give the same results.O8 Foliage from the leaves’ size and density:
foliage: {leaf_area_per_m3, leaf_width}works out what the leaves scatter and absorb in each band, and how far forward, from thin rigid discs turned every way alike; checked against the disc’s closed forms at both ends and a wall’s boundary-layer absorption, the late rays through such leaves against the particle simulation, and a wood of them compared with ISO 9613-2’s dense-foliage table. Presets and measured rates stay as they were.O9 Terrain and buildings from a map:
ground.terrain(heights on a grid from an Esri ASCII grid, x y z points or the scene,ground.map_origin) andgeometry.buildings_file(GeoJSON footprints raised to a property’s height, named, some left out); buildings, barriers, vegetation and zones stand on the terrain. The early sound reflects off the mean plane of the ground’s profile (the standards’ ground reflection) and bends over the crests the line of sight cuts through, one UTD wedge per crest with both slopes’ reflections in its terms, the equivalent edge and thick-screen factor for a rounded top, and the nearest crest’s wave in sight, so the level is continuous across the edge of a hill’s shadow; the late rays bounce off the triangles. A ridge is within 0.1 dB of the exact wedge in every band, a flat terrain and a slope give the flat ground’s and the image’s paths, a smooth hill over soft ground lets 1.4 to 1.8 dB more past than the thin screen (its crest is a wedge open wider than a half-plane), over rigid ground the flanks’ reflections add. Scenes without a terrain are unchanged.O10 Outdoor scenes in the editor: File > New outdoor scene (a meadow to build on), buildings, barriers, vegetation (a belt, a hedge or a tree) and ground zones drawn in the plan from the Geometry form, the ground’s extent set there, a terrain from a file and the buildings of a map (asking which properties hold their heights and names), each with a button to remove it; the plan shows the vegetation and the zones. Saved scenes refer to their terrain and map by relative paths.
O11 A fast field program and the 1995 benchmark cases:
validation/ffp.py, the exact solution through layers of constant sound speed, within 0.001 dB of the exact ground solution in still air, andvalidation/benchmark_1995.py, the benchmark’s still air and 0.1/s gradients down and up, and the same stopping 100 m up, at 10, 100 and 1000 Hz to 10 km through the engine’s parabolic equation and as whole scenes per band. With the gradient stopping 100 m up, as the weather’s do, the parabolic equation is within 0.35 dB of the program to 1 km downwind and within 3 dB at 10 km at 100 Hz (11 dB at 1 kHz, 47 dB down), and parts from it beyond 1 to 2 km with the benchmark’s gradient going on to a kilometre up (O21’s wide-angle parabolic equation carries that); whole scenes are within 0.11 dB per band to 1 km downwind in every case and upwind with the benchmark’s gradient, 1.5 dB at 1 km upwind with the gradient stopping 100 m up, 44 dB down. The comparison found and fixed a notch at the join of the parabolic equation’s bands and the rays’ (see the changelog).O12 Outdoor scenes on the GPU: the GPU traces the rays bent by wind and temperature through the same chords as the CPU (ray for ray: the same turns and the same curve to the millimetre, in single precision without trigonometry), the rays through vegetation (foliage and stems, up to 16 media) and the open sky, and steps the wave solver’s grid with its speed of sound per cell;
simulation.late.deviceandsimulation.wave.device: gpu,bothandautotake outdoor scenes now. What reaches each listener in each band, and when, agrees within the noise of the ray count (validation/outdoor_gpu.py).Bent rays and vegetation on the GPU on Direct3D 12 (Windows): the shader compiler’s (FXC’s) optimizer got the vegetation wrong, so it goes without it on Windows; and Windows CI’s software GPU (WARP) lost what the bent rays’ code wrote after returning early from a function, which it no longer does. Each now agrees with the CPU there as on other GPUs (see Validation).
Bent rays through vegetation on the GPU on Direct3D 12: FXC takes more than an hour over the tracer with both, so the CPU traces a scene with both there; the newer compiler (DXC) hung over the tracer’s shaders in Windows CI.
O13 Eddies weaker with height:
turbulence.height_m(the height the structure parameters hold at) andfalloff(2/3 by default, the surface layer’s; up to 4/3 in strong sunshine) make the eddies weaken with height from that height up, no nearer the ground than their outer scale, in the coherence of each pair of paths (Cn² at their mean height), the scattering into shadows and the shaking of the refraction; on the command line’sinspectand in the editor. Checked against the parabolic equation through eddies strengthening right down to the ground (validation/turbulence.py --profile). Withoutheight_mnothing changes.O14 Grounds of extended reaction:
impedance.reaction: extendedgives each reflection the impedance of its own angle (sound travelling sideways inside the ground, Attenborough’s models), which tells for soft snow, moss, forest floors and thin layers; the early paths at each reflection’s angle, the parabolic equation at grazing incidence, the wave solver locally. Checked against the fast field program with the ground’s reflection at every angle (validation/ground_reflection.py --extended). Not withtwo_parameter;localby default, as before.O15 Weather that changes over the scene:
atmosphere.zones, polygons in the plan with weather of their own (the lee of buildings, a lake under an inversion, a warm car park), taking what they leave out from the scene’s; the parabolic equation and the wave solver go through them, the early paths over the mean profile along their way and the late rays through the mean over the scene. Checked against a Crank-Nicholson parabolic equation marched through the zones (validation/weather_zones.py); drawn and set in the editor. Without zones nothing changes.O16 Levels over a long term:
long_termruns the scene in weather classes with their shares of the day, the evening and the night, and gives each period’s level, the long-term level and L_den (or L_R) at every listener and map point. Checked against the formulas and ISO 9613-2’s meteorological correction (validation/long_term.py); set in the editor. Without it nothing changes.O17 Moving sources and Doppler:
walk.dopplerplays each early path through a delay that changes as the source or the listener goes, andpass_bygives L_Amax, the sound exposure level and L_Aeq (walkthrough --levels). Checked against the exact field of a passing source over a rigid ground (validation/doppler.py). Without it nothing changes.O18 Moving air in the wave solver:
simulation.wave.wind: movingcarries the sound along with the wind before each step, so sound a building sends back goes against the wind. Checked against the exact field in a uniform wind, a façade’s echo against ray theory and the parabolic equation in thewindypreset (validation/wave_wind.py). Runs on the CPU; the wind’s shear term on the air’s motion stays out. Its wind flows round the model’s solids as an ideal fluid’s (no wake, no eddies), checked against the exact flow past a sphere. The air of the grid’s absorbing layer is still, and its border sends back about half the wind’s Mach number of the sound that leaves. Without it nothing changes.The wave solver’s room for sound along its matched layer only on the sides that need it: a side the model’s own surfaces close gets none, which saves a hall open at one end 1.8 to 2.0 times the cells (see Open models under Wave-based solver).
The wave solver’s source and listeners read between the cell centres to a higher order: a cubic through four cells a side, where two a side 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; see Free field under Wave-based solver).
The wave solver’s walls where the model has them (
simulation.wave.walls: fitted): the model’s faces cut exactly through the cells, on the standard stencil, indoors. A box’s modes come within 0.05 % of the exact ones along the grid and 0.38 % turned by up to 14 degrees, where walls on cell faces are up to 5.2 % and 4.2 % off; steeper walls, concave edges, free edges and absorbing walls have the limits under Walls in Wave-based solver. Without it nothing changes.Fitted walls outdoors: in wind and temperature each cell has its own speed of sound, so every interior cell’s coefficients must first sum to at most 2 by construction, and the slow decay of open rooms with such a speed, seen with fitted walls, be traced.
Fitted walls that keep a surface up to its free edge (left out now in the cell the edge ends in), share a piece of air that holds no cell centre among its neighbours, and close cracks and T-junctions (refused now); see Known issues.
O19 Whole cities: edge diffraction skips the pairs of an edge and a wall or roof that no path can use before it traces a ray, on every core, with the same results: 900 buildings simulate in about 2 s on four cores instead of 81 s, and 3,600 in about 22 s, where 1,600 ran out of memory (
validation/city_speed.py); with the reflections that fade at edges of 4.0.0, about 17 and 56 s.O21 A wide-angle parabolic equation:
simulation.early.parabolic_equation: wide_anglegives the bands the parabolic equation gives from the split-step Padé method (Collins’ RAM: ten terms, steps of three wavelengths, fourth-order differences in height, its own starting field), for gradients that bend sound through large angles over several kilometres. On the 1995 benchmark it is within 0.11 dB of the fast field program from 10 Hz to 1 kHz and to 10 km wherever the sound is within 40 dB of free field, where the narrow-angle one strays by 3 to 20 dB; in the weather presets within 0.03 dB on average at the bands’ centres (the narrow-angle one 0.10). Three to seven times the time over open ground; without it nothing changes.O22 Barriers and buildings in wind: behind them 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, checked against a full-wave solution and a parabolic equation started from a rigid barrier’s exact field (
validation/barrier_wind.py); the paths round a barrier’s end or a building’s corner take the open ground’s effect along their own way, for listeners in the open too. Still air is as before.Barriers in wind, next: the barrier’s own wind; the sound the eddies scatter over a barrier’s top into its shadow; the barrier’s runs at 4 and 8 kHz and the open ground’s shadow above its equation’s bands made to agree in a deep upwind shadow, where they are 12 to 40 dB apart and the level steps across the line of a barrier’s end below the line of sight by up to 12.8 dB; the paths round ends checked against a reference in three dimensions; the wide-angle equation behind barriers upwind in strong wind, where the narrow-angle one comes out low; the sound going back and forth between two barriers, which a one-way equation leaves out; and runs at 4 and 8 kHz, and at steep points, that cost less (see Known issues).
The eddies’ decorrelation of a path round an edge against the path it stands for (the direct sound in still air, the open ground’s path in wind), which steps across an end’s shadow boundary by up to 0.95 dB in still air and 1.08 dB in wind (see Known issues).
The path round a barrier’s end and then over a roof’s edge, without which the level 1.5 m over a roof 45 m from a source 1 m up (and the two swapped) steps across the end’s shadow boundary by up to 2.79 dB in still air and the weather’s effect by up to 0.57 dB (see Known issues); in wind it would also need the lit side’s runs, which the routing in
engine/refracted.pydoes not give a path round a vertical and then a horizontal edge.A transition across a caustic for the rays’ bands over the open ground (4 and 8 kHz, and 1 and 2 kHz with
open_ground_frequencies: rays), a uniform (Airy) form for the two paths that merge there: past it they vanish and the level steps by up to 15.5 dB within a centimetre (up to 14.0 dB at 1 and 2 kHz withrays; see Known issues).The parabolic equation along a wall’s reflection in wind and temperature gradients, as over the open ground: the paths off façades and barriers are eigenrays in every band, with the rays’ ground effect up to 2 kHz and their steps at caustics (see Known issues).
An absorbing layer for the parabolic equation’s runs behind barriers and buildings that takes up the sound an obstacle’s top sends steeply up, and that a run’s other listeners do not move: there the listeners and map points that share a run share its layer, and the level changes with where it begins, by up to 2.4 dB behind a tall barrier close to the source in the checks (see Known issues).
The crossover from the parabolic equation’s bands to the rays’ (2.4 to 3.4 kHz, or 0.59 to 0.84 kHz with
open_ground_frequencies: raysshort of half the way to the shadow): out of the shadow the equation’s field goes through it, but the rays’ band beside it is 5 to 22 dB off the equation there in the weather presets, the paths round a barrier’s end below its line of sight still hold the weather’s effect at their bands’ early frequencies there (1.1 dB off denser runs 100 m downwind inwindy), the paths over a roof’s edge and behind barriers take one ratio of the equation’s band energies a band there, and in the shadow, where the bands above the equation deepen from the octave below 18 to 36 dB louder than its field, the bands keep their own samples (8.6 dB low in case 3 300 m away; see Known issues).The eddies downwind: their coherence factor, fitted in still air and in the shadow, leaves the 1 kHz band 0.3 to 1.8 dB from a parabolic equation through random eddies 100 to 300 m downwind, and with eddies a step of the rays at a caustic comes into the equation’s bands; and the echogram of the bands up to 500 Hz, which takes the coherence as a band’s mean (see Known issues).
Steep sound over the open ground in wind (as under a flyover) left to the rays at 1 and 2 kHz by a rule, where they match the equation at a fraction of its time: one studied (the vertical wavenumber changing by at most 3 % over one of its own lengths along the paths) was within 0.03 dB of a fast field program in the presets and under linear gradients and took 0.14 to 0.37 of the equation’s time on the helicopter flyover, but its error has no bound (0.12 dB past a 3 K step a centimetre above the source); until one holds,
open_ground_frequencies: raysis the user’s choice.The scattered sound of the reflections straight from the source found from each listener too, shared with the late rays by how densely each finds it: under an aircraft overhead a listener above a ground that scatters no longer gets bursts of it from the few rays that land nearby.
The same where the rays bend in wind or a temperature gradient, scatter in plants before they reflect, or reach the listeners as secondary sources: there the scattered sound near a listener still comes from the few rays that land nearby.
Sources:
Coherent arrays added up element by element near their listeners (
near_field: true): the direct sound, the reflections and the sound that reaches an edge, exact over each band and the pattern far away.Sources and listeners in groups, and left out of a run without being removed (
group,enabled: false); in the editor a tick and an eye per source, listener and group.A measured directivity table takes the directions equally near a point of its grid alike, and a direction given more than once as one.
Examples:
A concert in a stadium with a sound system: main line arrays, subwoofers, fills, delay towers and delays timed to them, with the stage’s monitors (
stadium_concert, whose sound system plays alone with its stage and monitors left out); the stage alone with the singer’s microphone’s gain before feedback (concert_stage); every array added up box by box near its listeners; the system tuned at the mix position, as a system engineer tunes it; and the editor’s examples in groups.Demos with their scripts: a choir and an organ in the church, a band with its sound system in the stadium (one of three songs) and a song carried across a lake by a lawn concert’s line arrays and subwoofers, in the evening’s weather and in still air, and the orchestra in the concert hall, its four seats toured through three movements; each downloads its recordings, mixes them at every listener and makes a video that tours them. The stadium’s pictures show where the crowd stands and sits, the empty seats, and how much each takes up.
Everything a demo plays in its example’s scene, left out of the example’s own simulations until ticked, so the editor shows it: the stadium’s band and the singer’s microphone, the church’s choir, organ pipes and toured places, the lakeside’s sound system and its listener behind the stage, the gallery’s place to look round from.
A walk through a gallery, the fifth demo and the first of speech: a guide talks while a visitor walks away and round a corner, then turns their head in front of them (the
galleryexample).One shared package behind the demo scripts (
acoustic_spaces.demos): a demo describes its steps and options and gets the command line, the recordings, the simulation, the mixes, the tour, the stills and the video, and runs from the editor as well as from a terminal.Free HRTF sets to download, with their licences, from the command line (
acoustic-spaces hrtf) and from the editor.A demo of any scene (
acoustic-spaces demo): each source plays the WAV file it names, at its sound power, and every listener is heard in turn in a video drawn from the scene; every demo’s stills say how the sound was made and the command that makes it again.A recording of the user’s for any source of any demo (
--input SOURCE=WAV), at the level of the feed it replaces, and a table of the sources in Run demo…A song of the user’s own in the stadium and lakeside demos (
--music,--bpm,--credit; the stadium’s stems with--stem NAME=WAVand its first downbeat), nothing downloaded, with a file field, a number and a table of the stems in Run demo…A recording for a scene’s sound system in the demo of any scene: the scene’s
recording(a file of the user’s, or a free song or talk downloaded the first time) and each source’schannelof it (left, right or mono), on the scene’s and the source’s pages in the editor; every example with a stage, a sound system or a talker plays one out of the box, its talkers at the normal vocal effort of IEC 60268-16.
Movement:
Walks along timed keyframes (position and facing at each time, waits and turns on the spot), walks that can be switched off, and every mover at once on one clock in one simulation (
walkthrough --all), heard at every listener, with a video of them all.Editor: the Movement panel, with the keyframes of a source or listener in a table (no speed to set: it follows from the times), a timeline that can be made longer to add keyframes past the end, a preview of every mover in the 3D view and the plan, draggable keyframe handles, and the movement rendered in the background.
One scene still or moving: walks turned on and off rather than copies of a scene with and without them; the demo of any scene plays the walks that are on, with a moving picture, and
--movingpicks them for a run; the Run demo window ticks what moves and opens the Movement panel.A car passing by with Doppler: a car in the
roadsideandstreetexamples with CNOSSOS-EU’s sound power, a pass-by demo for both with a sound made to that power and the level over time, the demo of any scene playing one walk with Doppler, and the Movement panel’s Doppler tick.Editing a car’s walk as it is: a walk along points edited in the Movement panel (points, one speed), keyframes at a steady speed accepted by the pass-by demo, Render movement… heard binaurally and with Doppler, and units for every number in the panel and in Run demo, with the pass’s length, speed and duration.
Faster walks outdoors: the places of a walking source share the parabolic equation’s runs, each listener over the open ground getting what a run of its own place would give it.
The sound a source sends each way, along a walk with Doppler: each early path carries what is sent the way it leaves the source (
acoustic_spaces.emission).A helicopter’s rotors: their tones from the blades’ steady loads, checked against Gutin’s hover formula and Farassat’s formulation in time, and Pegg’s broadband noise, checked against his worked example (
acoustic_spaces.rotorcraft).A helicopter’s broadband noise from each blade section’s own noise (Brooks, Pope and Marcolini’s airfoil self-noise), checked against its authors’ sample outputs, or from Pegg’s method, chosen per helicopter.
A helicopter’s sound within its certified level: the AS350 B3’s overflight comes out 3.4 dB over with Pegg’s broadband noise and 4.5 dB under with the blade sections’ own, which leaves out the blades cutting through the wake of those ahead and through turbulence.
A helicopter flying over a meadow: the
helicopter_flyoverexample and its flyover demo, the rotors trimmed at the helicopter’s airspeed against the wind.A helicopter crossing over the stadium concert:
helicopter_demo.py, its pass added to the concert’s mixes at their simulated levels, 150 m up halfway through the song, or 330 m as the rules of the air ask over a crowd without leave. The helicopter and its walk are in the stadium’s scene, left out of the concert’s simulations, to see and change in the editor.Trucks, buses, motorcycles and a police car with its siren on, in the pass-by demos as the car is (
--vehicle): CNOSSOS-EU’s sound power of their category, and the siren’s level and directivity as measured, checked against California’s and the Swiss rules for sirens; the police car’s own sound (CNOSSOS-EU’s car) along with its siren, or the siren alone.A fire engine in the street: a medium heavy vehicle (CNOSSOS-EU’s category 2) with its siren’s two tones as loud as the Swiss rule allows on a vehicle and an air horn in blasts as loud as the rule for vehicles’ horns allows on its bench, both checked against those rules; the vehicle, the siren and the horn heard together, or without the horn or the vehicle.
Trains and trams by Schall 03: the
railwayexample, its demo running a regional train, an S-Bahn, an intercity or a freight train past as long as each is, and a tram in the street; the engine’s pass checked against Schall 03’s own sum.CNOSSOS-EU’s railway noise for the trains, to choose instead of Schall 03: second to it, as the Directive’s default transfer functions give about 6 dB(A) more rolling noise than TWINS and leave trams poorly covered.
Construction plant and reversing alarms: road works in the street (a breaker on an excavator, a generator and a dumper reversing with its alarm on) from Defra’s noise database for construction sites, and a demo that plays them; checked against Defra’s rows, the EU’s limits for outdoor machines and SAE J994.
Car horns: the street’s car with a pair of disc horns (
passby_demo.py --car car_and_horn), as loud as a maker’s approved to UN Regulation 28, checked against the rule for vehicles’ horns; a horn’s levels in the captions follow its blasts.A train’s warning horn (
passby_demo.py --horn blastson the railway): the TSI’s two notes in turn (or together,--horn chord), as loud as EN 15153-2 allows 25 m ahead, the measuring site’s ground taken out harmonic by harmonic.A train horn’s directivity measured, mounted on a train’s front (Toward et al. found one 1 dB down 30 degrees off its axis, where the siren the horns take is 3.6 dB down).
Aircraft noise at receivers by ECAC Doc 29: one movement’s SEL and LAmax from the NPD curves along the flight path’s segments, with Volume 3’s three hypothetical aircraft, checked against all of Volume 3’s reference levels, grids and segment results.
Doc 29’s flight performance (Volume 2 Appendix B): profiles from procedural steps at any aerodrome, checked against all of Volume 3 Part 2’s reference profiles; tables read from a copy of the ANP database of one’s own.
Doc 29’s bank angles in turns: from the ground track’s turn radii (Volume 2 Appendix B5, eq. B-14) for the engines’ installation correction, and their effect on the climb in the performance calculation.
Aircraft taking off and landing in a scene, heard in a demo, its levels checked against Doc 29’s: the airport example’s twin jet, flown by Doc 29’s Appendix B.
Houses under the airport example’s flights, and an aircraft flying over at cruise.
The L_Amax of a pass with fast time weighting, and with slow time weighting in the captions of vehicles other than the aircraft (whose captions give it), as measured, beside the highest quasi-static level the captions give now.
Ground-borne vibration from trains, trams and construction plant, and the rumble it makes in a building’s rooms, by the US FTA’s general method: overall levels from its curves and tables against its limits and damage criteria, for the railway’s trains and the street’s tram and road works; checked against the manual’s figure, worked example and tables.
The pass-by demos and the road works heard indoors by listeners in front of a building’s wall: the train, the tram or the works through a closed window (EN ISO 12354-3) and the rumble of the ground’s vibration; at every listener, what is felt, in the captions and as a bass shaker’s track.
Ground-borne vibration by the FTA’s detailed method: band by band, from a train’s force density and a site’s transfer mobility (read off the manual’s measured curves or measured), and the floors and walls radiating it into the rooms in the engine, so the rumble is heard in a building’s simulated rooms, not only reckoned.
Ground-borne vibration by the ISO and DIN norms, beside the FTA’s manual, which is US guidance: rail systems’ vibration predicted as ISO 14837-1 guides, people in buildings by ISO 2631-2 and DIN 4150-2, and damage to buildings by DIN 4150-3.
Movement: a path drawn in the plan click by click.
Editor:
A simpler Run demo window: the options that change what a demo plays, in words, and the step, the paths, what fine-tunes a run and the command under Advanced.
An example’s scripts that are no demos (the singer’s microphone’s feedback, the loudspeakers’ alignment) run from the Run demo window.
The Run demo window says what a run downloads and how much, above Run; the orchestra demo downloads its tracks itself, as the other demos do their recordings.
The listeners a demo plays, picked for a run:
--listenersin every demo with several listeners, and a Listeners list to tick in the Run demo window.The Run demo window keeps the options changed in it for the next time, per demo and, for what names something in a scene, per scene file, with Defaults to put them back.
The background task in the status bar says what runs, with Cancel beside it; messages that can be shown or hidden by level, filtered by item or words, and cleared, every other one shaded.
Markers sized from the spacing of the sources and listeners, smaller where they would overlap, and arrows that stop short of a marker ahead.
The wind drawn outdoors, in the 3D view and the plan: arrows the way it blows, as long as it is strong, the weather zones with their winds, and masts with the wind’s profile up them; eyes for both in the scene tree, and an example to see them (
lakeside_concert).The particle export to Blender asks how much of the sound, how slowly and how many particles, outdoors at first enough to reach the farthest listener.
Simulation > Run demo…: the demos of the open scene (This scene, and the example’s own for an example’s scene), its options at the command line’s defaults, run in the background with its messages in the panel; the source page’s Plays row picks the WAV file each source plays, and the window’s table of sources the file each plays instead in that run (
--input).Download HRTFs…: the free sets with their licences, fetched into the user’s data folder and put in the scene’s binaural output.
The Scene YAML panel: the scene as coloured YAML text that follows every edit, and edited text applied with the checks every edit goes through.
The views follow a run: the sources and listeners its step works on ringed, a walking source’s places shown as they are simulated, and a demo’s video followed moment by moment as it is encoded.
Running on other computers, in steps:
Packs (
acoustic-spaces pack, File > Pack for another computer): a scene and every file it names in one zip, which every command that reads a scene takes, and a container image that runs them headless on a GPU through Vulkan.acoustic-spaces remote: send a pack to a server over SSH or to Hugging Face Jobs, follow its progress, cancel it and bring the results back.Run remotely in the editor, with the results opened from the folder that comes back.
Version 4.0.1:
The archway’s 1.4 dB, where the reflection off the jamb is found: in the bands that add in energy the terms are kept as waves, whose energy is never below nothing, and the events a path gets past split like those that stop it, so the level steps there by at most 0.11 dB from 250 Hz to 1 kHz.
A term past the foot of its edge, where the edge stands on the plane of the next reflection, goes on as the term of the edge’s image: behind the archway’s wall the level steps by at most 0.08 dB from 250 Hz to 1 kHz, and no longer by 0.63 dB at 125 Hz at y = 6.32 m.
Reflection phases from an admittance fitted to a material’s absorption in every band, so that the level at 63 and 125 Hz no longer depends on how many bands add in pressure.
Rooms with a curved surface of many facets, as BRAS CR2 and CR3, run at their defaults instead of running out of memory, in octave and in third-octave bands.
The validation report’s measured rooms (BRAS) and its cross-check with pyroomacoustics run again with this version, with their run times and memory.
The edge fade walked in CI: the
archway’s whole line behind its wall, thegalleryround its corner and thestreet’s garden round its four largest steps (tests/test_walks.py), withvalidation/walks.pyfor the whole walks. The edge diffraction and the fade checked against the exact solution past a thin screen’s corner, behind a barrier on a hard ground withfirst_order(one_sideandboth_sides) and behind a thick barrier withsecond_order.
Known issues found by the review of 25 October 2026 (see Validation), in the order they come next:
Paths through three partitions or more, which neither part counted: the late stage counts those that reflect, and the early stage leaves them to it.
Packs that keep an earlier one when writing fails, and result file names that can’t be the same for two names.
Vertices that no face uses dropped, coordinates that are not finite refused in every format, and a solid standing on a surface (a block on the floor) found to be solid from several of its faces.
Sources and listeners inside a closed solid reported, and
transmission: 0as a surface that lets nothing through.The late part’s rain kept from going out of the rooms and back in, as the rays are.
G against a free-field response that starts at its first sample: the free-field reference keeps what its band filters spread before it.
The first milliseconds of a rendered response, where the band filters’ spread before the first sample is lost: filters that don’t spread before an arrival.
No path arriving before it leaves the source, and errors in the native code’s threads raised in Python.
The echogram’s pairs of paths from 1 kHz up without adding energy where they cancel.
A warning for coefficients outside 0 to 1, and the wave solver’s grid checked before the weather’s fields and the first one let go round an open model.
The music echo criterion through one band-pass over its two octaves, as Dietsch and Kraak have it.
Limits on the size of a run: its memory estimated before it starts, for the computer’s cores and devices, and a run past the computer’s memory refused (threads above the cores run as the cores).
Results written whole or not at all, a response at a time, and
simulate’s folder checked before its run.The mixing time from the listener’s room (the detector stays sized for all the rooms, which the rays spread over).
The real-time convolver writing each block into buffers made before playback starts.
A warning when the wave solver’s grid is coarser than the default, with how slow its shortest waves travel, and that in each run’s details.
The wave solver’s buffers on a GPU all checked against what it binds, and its recording at the listeners kept once.
The late part’s last milliseconds with the whole chords of the rays through the listeners’ spheres and the map’s cells.
Re-weighting for other absorption (the editor’s results after an absorption change) turns the phases of the early reflections where they add in pressure, as tracing again has them.
The wave solver’s highest modes kept inside its stability limit in single precision, with both stencils.
The review’s lower-severity findings, each checked: faces given twice dropped, sources and listeners on a surface refused and listeners next to one warned of,
validate --strict, YAML keys written twice refused, packs unpacked only with room for them and each file checked before it takes its name, the late part’s smoothing kept after 50 ms, the tracer’s reflection counts in 32 bits, andfrom_impulse_responseandwave_responseschecking what they are given. The others were not borne out, or only a direct caller of the native module can reach them.CI: type checks, sanitizers for the native code and the tests’ coverage.
Type checks passing in the modules that did not when they began (listed in
pyproject.toml), a few at a time.Collada polygons with their holes.
The wave solver’s walls through planes of its cell centres, and sources and listeners on surfaces.
The wave solver’s surfaces a hair off its cell centres (see Validation, Known issues).
Vector-based scattering with a lobe centred on the specular direction, so the late part carries the specular share of the orders after the first as the exact paths do, checked against the BRAS rooms’ decay.
Version 4.1.0 (and, in the sections above, the helicopters, the aircraft, the road vehicles, the trains, the road works, the horns, the ground’s vibration, the pass-by demos and the road works heard indoors and the faster walks outdoors under Movement, the demos’ elements in their scenes and a recording for a scene’s sound system under Examples, and the Run demo window, the listeners a demo plays and the Scene YAML panel under Editor):
Uncertainty: each parameter’s standard error, how much it would change between runs with other seeds, from the late rays traced in batches (a jackknife over them), in
parameters.json,simulate’s output and the editor’s results; checked against the spread over seeds.EDT’s line fit always from the onset, however the energy is summed.
The path over a barrier’s top with the reflection off the ground behind it, kept where the reflection before the barrier is found: behind a barrier on a hard ground the level follows the exact one within 0.05 dB with
first_orderandboth_sides, instead of stepping by 0.35 dB at 63 Hz. Checked against BRAS: in the seminar room 12 terms over its 10 pairs that were left out count now, which moves its early level by up to 0.8 dB at 63 Hz.Models cleaned where an edge runs past a corner at which three faces meet: BRAS’s auditorium (CR4) is closed.
More measured rooms: BRAS’s coupled rooms (CR1, the door open by 4.1°) and auditorium (CR4) beside the seminar room and the hall, G from seat to seat in all four with the measured level’s 3 dB question reported, IACC of FABIAN’s binaural responses in three (
validation/bras_binaural.py), and each run’s largest standard error.Random scattering traced band by band where the bands scatter differently, as vector-based scattering is: with one ray for all bands, re-weighted band by band, its decays in BRAS were noisy, and in the coupled rooms far too short.
Paths over edges that arrive together rendered with phases of their own, so they render their energy on average over seeds instead of adding in amplitude; the echogram’s 63 and 125 Hz reading no longer takes their interference a second time from the band above.
Listeners as microphones: a first-order pattern, or a coincident pair (XY, Blumlein, MS) in two channels, that the mono output picks up through, aimed where the listener faces; chosen in the editor’s listener form and drawn as balloons in the 3D view.
Microphone patterns that change with frequency: one per octave or third-octave band, for a capsule or a pair’s capsules, rendered exactly from the same paths.
Heavy vehicles, buses and motorcycles in the pass-by demos with CNOSSOS-EU’s sound power of their category (a truck in the roadside, a bus and a motorcycle in the street), and a police car with its siren on; the car’s sound power by the 2021 amendment of CNOSSOS-EU’s table, 2.5 dB(A) more at 50 km/h.
Version 4.2.0 (in the sections above: the fade past a thin screen’s free ends under Edge diffraction, the error estimates under Parameters, barriers and buildings in wind (O22) and the scattered sound found from each listener under Outdoor acoustics, and the Run demo window’s remembered options and the views following a run under Editor):
BRAS’s coupled rooms with the door open by 30.4°, whose model had no inside (the door leaf ran 24 mm into the wall at its hinge): the conversion joins the leaf to the wall and floor.
Version 4.2.1:
A user manual that leads through the documentation, online with the documentation and as one PDF.
ISO 12999-2’s octave-band values (Table 2) for the built-in materials’ spread from 250 Hz to 4 kHz.
Next:
A microphone’s measured on-axis response and measured patterns (files), as sources have.
Spaced pairs whose late parts correlate as real capsules’ do at low frequencies.
ISO 3382-3’s quantities for open-plan offices on a line of positions, with its normal speech spectrum and the background noise: the spatial decay rate D2,S and Lp,A,S,4m fitted to the speech level against distance, and the distraction, privacy and comfort distances from the STI and the level along the line.
The package’s sums in a fixed order (Simulation) as fast as the linear-algebra library’s, with the same bits: wider vector instructions chosen when it runs, and less to do for a small product. A rotor’s
bpmbroadband noise takes 23 % longer than it did with the library’s sums.A walk’s places rendered on the run’s threads, as a run’s files are: they are rendered one after the other.
A response’s echogram worked out once for all that read it: a run works it out for the parameters, for the speech results, for the standard errors and, outdoors, for the level against the free field. And the early paths’ signals on the run’s threads.
The command line starting without the engine’s imports, for the commands that do not need them.
Less memory while a response is rendered: its early paths placed a part at a time, and its bands combined in place.