Skip to content

Room acoustic parameters (ISO 3382-1/2)

Standards: ISO 3382 Key references: Kuttruff 2016 Schroeder 1965 Hak et al. 2012

Verified by 23 conformance checks against ISO 3382-1:2009 and ISO 3382-2:2008: 23 in Room & building acoustics.

Room acoustics starts from one measurement: the impulse response (IR) between a source and a receiver. Filter it into bands and integrate it, and it yields reverberation time, clarity and centre time: everything ISO 3382-1/2 asks about the sound field inside a single room. This page turns a measured IR into those parameters: the Schroeder decay curve, the evaluation windows, the validity flags and the accredited fiche. Acquiring the IR itself (sweeps, MLS, where to measure) lives in Measuring the Room Impulse Response, and the open-plan speech metrics of ISO 3382-3 in Open-Plan Office Acoustics. Two separate decay measurements in a reverberation room, one with the room empty and one with the specimen installed, yield a material’s sound absorption; that method (ISO 354) lives in Sound Absorption Measurement and Rating. For sound insulation between spaces (the same IR measured either side of a partition) see the companion Field Insulation Measurement (ISO 16283) guide.

How do I compute EDT, T20 and T30 from an impulse response in Python?

Section titled “How do I compute EDT, T20 and T30 from an impulse response in Python?”

Pass the impulse response and its sample rate to room.room_parameters(ir, fs). It splits the response into octave bands (fraction=3 gives one-third octaves), integrates each band backwards into its Schroeder decay curve and fits the evaluation range of each parameter, so res.edt, res.t20 and res.t30 hold one decay time in seconds per band, with the clarity c50 and c80, the definition d50 and the centre time ts beside them. res.t30_valid says, band by band, whether the noise floor left the fit enough decay range. On the stand-in response below, the octave values of res.t30 average to 1.21 s. What it means when EDT, T20 and T30 disagree is in Reading EDT, T20 and T30 against each other.

import numpy as np
from phonometry import room
# ir: your measured impulse response, sampled at fs hertz. Two seconds of noise
# decaying with a reverberation time of 1.2 s stand in for it here.
fs = 48000
t = np.arange(2 * fs) / fs
ir = np.random.default_rng(1).standard_normal(t.size) * np.exp(-3 * np.log(10) / 1.2 * t)
res = room.room_parameters(ir, fs)
print(res.frequencies.round()) # [ 126. 251. 501. 1000. 1995. 3981.]
print(res.edt.round(2)) # [1.16 1.29 1.16 1.12 1.19 1.17] s
print(res.t20.round(2)) # [1.18 1.24 1.21 1.25 1.24 1.2 ] s
print(res.t30.round(2)) # [1.2 1.19 1.23 1.21 1.21 1.21] s
print(res.t30_valid.all()) # True
print(round(float(res.t30.mean()), 2)) # 1.21 s

Decay analysis and room parameters (ISO 3382-1/2)

Section titled “Decay analysis and room parameters (ISO 3382-1/2)”

The IR is filtered into octave (or one-third-octave) bands and each band is turned into a decay curve by Schroeder backward integration of the squared IR:

Integrating backwards removes the fluctuation that plagues a raw squared IR and yields a smooth curve whose slope is the decay rate. Background noise would make level off, so integration is truncated where the fitted decay line crosses the noise floor and the missing tail is compensated assuming an exponential decay.

Reverberation times come from a least-squares line fit over an evaluation range, extrapolated to a full 60 dB drop, : EDT over 0 to −10 dB (perceived reverberance), T20 over −5 to −25 dB and T30 over −5 to −35 dB. Energy splits at an early/late boundary give clarity and definition,

with ms → C50 (speech) and ms → C80 (music), plus the centre time . Each parameter has a just-noticeable difference (ISO 3382-1 Table A.1: EDT 5 %, C80 1 dB, D50 0.05, Ts 10 ms) that sets how precisely it is worth reporting.

Where is. All three energy integrals start at the arrival of the direct sound, not at the start of the file. ISO 3382-1 A.3.4 places that point where the response first rises significantly above the background while still being more than 20 dB below its maximum, and room_parameters applies that trigger per band before integrating, so any propagation and system delay carried by the recording is removed automatically. A measured IR needs no manual trimming. It matters asymmetrically. The reverberation times use only the slope of the decay and are indifferent to where it starts, but C50, C80 and integrate from , so an onset placed one millisecond late throws away part of the direct sound and can move C50 and C80 by several decibels and by tens of milliseconds, while leaving T20 and T30 almost untouched. The error is invisible in the number everyone checks. The detector deliberately errs early, which is the harmless direction. Two working rules follow: never trim an IR by hand to remove its pre-delay, and be suspicious of C50/C80 from a recording whose direct sound is soft or has been pre-ringed by external processing, because the trigger can then fire late.

The tail energy of the squared impulse response fills from the end while the backward integral advances toward t = 0, and the Schroeder decay curve emerges on a companion axis, ending with the T20 and T30 regression lines.

Download the animation (WebM)

The tail energy of the squared impulse response fills from the end while the backward integral advances toward t = 0, and the Schroeder decay curve emerges on a companion axis, ending with the T20 and T30 regression lines.

Download the animation (WebM)

Squared impulse response with its Schroeder backward-integrated decay curve, and the EDT, T20 and T30 regression windows markedSquared impulse response with its Schroeder backward-integrated decay curve, and the EDT, T20 and T30 regression windows marked

The jagged squared IR (grey) integrates to the smooth Schroeder curve (blue); the EDT, T20 and T30 windows are fitted on that curve and each extrapolated to a 60 dB decay.

import numpy as np
from phonometry import room
fs = 48000
# Single-slope decay with T = 1 s: p^2 = exp(-13.8155 t) (60/ln(10)/13.8155 = 1)
t = np.arange(fs) / fs
# ir: measured room impulse response; a synthetic single-slope decay stands in here.
ir = np.concatenate([np.zeros(10), np.exp(-13.8155 * t / 2.0)])
decay = room.decay_curve(ir, fs) # Schroeder curve (0 dB at t = 0)
time, level = decay.times, decay.levels
res = room.room_parameters(ir, fs, limits=None) # broadband single band
print(round(float(res.t30[0]), 2)) # 1.0 s
print(round(float(res.c80[0]), 2)) # 3.05 dB
print(round(float(res.d50[0]), 3)) # 0.499
print(round(float(res.ts[0]) * 1000, 0)) # 72 ms
# Octave bands 125 Hz - 4 kHz (ISO 3382-1 default); use fraction=3 for thirds
octaves = room.room_parameters(ir, fs)
print(octaves.frequencies) # ~[126, 251, 501, 1000, 1995, 3981]
print(octaves.t30_valid) # per-band dynamic-range flags
octaves.plot() # per-band EDT/T20/T30 + C50/C80 bars (needs matplotlib)
room.decay_curve(ir, fs).plot() # Schroeder decay with EDT/T20/T30 fit overlays
Show the code for this figure
import matplotlib.pyplot as plt
import numpy as np
from phonometry import room
fs = 48000
# Single-slope decay with T = 1 s: p^2 = exp(-13.8155 t) (60/ln(10)/13.8155 = 1)
t = np.arange(fs) / fs
# ir: measured room impulse response; a synthetic single-slope decay stands in here.
ir = np.concatenate([np.zeros(10), np.exp(-13.8155 * t / 2.0)])
decay = room.decay_curve(ir, fs) # Schroeder curve (0 dB at t = 0)
time, level = decay.times, decay.levels
# One line — Schroeder decay with the EDT/T20/T30 straight-line fits:
decay = room.decay_curve(ir, fs) # a DecayCurve: read decay.times and decay.levels
decay.plot()
plt.show()
# By hand, the decay is just the Schroeder curve; mark the evaluation levels:
fig, ax = plt.subplots()
ax.plot(time, level, color="#1f77b4", label="Schroeder decay")
for db in (-5.0, -25.0, -35.0): # T20 / T30 evaluation-window edges
ax.axhline(db, ls=":", alpha=0.4)
ax.set_xlabel("Time [s]")
ax.set_ylabel("Level re steady state [dB]")
ax.set_ylim(top=3.0)
ax.legend()
plt.show()

For this single-slope decay EDT, T20 and T30 all return ≈ 1.0 s, and the energy parameters match their closed forms (C80 = 3.05 dB, D50 = 0.499, Ts = 72 ms). A real room has a steeper early slope, so EDT < T30.

What the energy parameters mean. C50 and D50 are the same measurement in two units, tied by ; the printed pair above satisfies it, since = 0.499 gives = −0.02 dB, so a speech space aiming at is aiming at dB, and there is no point quoting both as if they were independent evidence. For music, C80 in unoccupied concert and multi-purpose halls up to 25 000 m³ typically runs between −5 dB and +5 dB (the “typical range” column of ISO 3382-1 Table A.1), with negative values reading as blended or reverberant and values above about +3 dB as dry; in the same halls runs between 60 ms and 260 ms and moves inversely with clarity, and between 0.3 and 0.7. Two cautions go with those ranges. They are single-position values in unoccupied halls, and an audience shortens the decay and raises clarity, so an occupied hall is not described by them. And unlike T30, the energy parameters are early-energy quantities that vary strongly from seat to seat, so a single-position value means little: report the spatial mean with its spread (below). The JND column sets the reporting resolution in the other direction: a 0.5 dB difference in C80 between two designs is not a difference at all.

On a real, frequency-dependent decay the per-band .plot() is the working summary of the whole measurement: the decay times as grouped bars per octave (invalid bands hatched) over a second panel with C50 and C80.

ISO 3382 per-band parameters of a synthetic room impulse response: grouped EDT, T20 and T30 bars per octave band falling from about 1.4 s at 125 Hz to 0.7 s at 4 kHz, over a second panel where C50 and C80 rise with frequencyISO 3382 per-band parameters of a synthetic room impulse response: grouped EDT, T20 and T30 bars per octave band falling from about 1.4 s at 125 Hz to 0.7 s at 4 kHz, over a second panel where C50 and C80 rise with frequency

A room whose reverberation time falls from 1.4 s at 125 Hz to 0.7 s at 4 kHz, the typical signature of a furnished room whose absorption grows with frequency; clarity rises as the decay shortens.

Show the code for this figure
import matplotlib.pyplot as plt
import numpy as np
from scipy import signal
from phonometry import room
# A synthetic room IR with a frequency-dependent decay: octave-band noise
# carriers whose T60 falls from 1.4 s at 125 Hz to 0.7 s at 4 kHz.
fs = 48000
rng = np.random.default_rng(3382)
t = np.arange(int(1.6 * fs)) / fs
ir = np.zeros_like(t)
for fc, t60 in [(125.0, 1.4), (250.0, 1.25), (500.0, 1.1),
(1000.0, 1.0), (2000.0, 0.85), (4000.0, 0.7)]:
sos = signal.butter(4, [fc / np.sqrt(2), fc * np.sqrt(2)],
btype="bandpass", fs=fs, output="sos")
carrier = signal.sosfilt(sos, rng.standard_normal(t.size))
ir += carrier * np.exp(-3.0 * np.log(10.0) / t60 * t)
# One line: per-band EDT/T20/T30 bars + C50/C80 (needs matplotlib).
octaves = room.room_parameters(ir, fs)
octaves.plot()
plt.show()
# By hand: the T30 spectrum from the result's fields.
fig, ax = plt.subplots()
ax.bar(np.arange(octaves.t30.size), octaves.t30)
ax.set_xticks(np.arange(octaves.t30.size))
ax.set_xticklabels([f"{f:g}" for f in octaves.frequencies])
ax.set_xlabel("Octave-band centre frequency [Hz]")
ax.set_ylabel("T30 [s]")
plt.show()

Reading EDT, T20 and T30 against each other

Section titled “Reading EDT, T20 and T30 against each other”

The three times extrapolate the same 60 dB decay from different windows, so their disagreement carries information:

  • T20 ≈ T30 (curvature below 10 %): the decay is close to a single straight slope over both evaluation windows, which is consistent with (though not proof of) a diffuse field, and either time can stand for “the” reverberation time of the band.
  • T30 > T20 (curvature above 10 %): the decay sags, with late energy decaying more slowly than early energy. The usual causes are coupled volumes (an open door to a corridor or stairwell, a deep balcony or a stage house feeding energy back) and strongly uneven absorption that leaves one room axis reverberant. No single number describes such a decay: report both windows together with the curvature, and treat the statistical predictions with suspicion, because their diffuse-field assumption has visibly failed.
  • EDT far from T20/T30: EDT is fitted where the direct sound and the first reflections still dominate, so it varies from seat to seat while T30 barely moves. EDT below T30 means the position receives strong early energy (close to the source, under a reflector): the room sounds drier there than its T30 suggests, because perceived reverberance follows EDT. EDT above T30 at one seat points to an echo or a focusing surface concentrating late energy there.

The three cases are three shapes, and they are worth seeing side by side before you have to diagnose one.

Three panels sharing a level axis, each a Schroeder decay curve of a synthetic impulse response with its EDT, T20, T30 and curvature printed: a straight single-slope decay with curvature near zero, a sagging two-exponential decay with a visible knee and curvature well above ten per cent, and a decay preceded by a strong direct sound and one early reflection whose early decay time falls far below the T30Three panels sharing a level axis, each a Schroeder decay curve of a synthetic impulse response with its EDT, T20, T30 and curvature printed: a straight single-slope decay with curvature near zero, a sagging two-exponential decay with a visible knee and curvature well above ten per cent, and a decay preceded by a strong direct sound and one early reflection whose early decay time falls far below the T30

Three decays, one axis, all read by room_parameters. Left, a single exponential at = 1.0 s: EDT, T20 and T30 all return 1.00 s and the curvature is −0.1 %. Middle, a coupled volume: 98 % of the energy decaying at 0.6 s over a 2.5 s tail, which puts a visible knee in the Schroeder curve and returns EDT 0.72 s, T20 1.78 s, T30 2.18 s and a curvature of 23 %: no single number describes it. Right, the same 1.0 s tail preceded by a cluster of strong early reflections in the first 25 ms: T20 and T30 are untouched at 0.99 s and 1.00 s, the curvature stays at 0.5 %, and EDT collapses to 0.58 s: the seat sounds drier than its reverberation time says, and only EDT knows.

Show the code for this figure
import matplotlib.pyplot as plt
# `np`, `fs` and `room` are the imports of the decay block above.
rng = np.random.default_rng(3382)
span = np.arange(int(4.0 * fs)) / fs
def energy(t60): # the p^2 envelope of a T60 = t60 decay
return np.exp(-6.0 * np.log(10.0) / t60 * span)
noise = rng.standard_normal(span.size)
tail = noise * np.sqrt(energy(1.0))
early = tail.copy() # a cluster of strong early reflections
amplitude = np.sqrt(6.0 * float(np.sum(tail ** 2)) / 8.0)
for k, i in enumerate(np.linspace(0, int(0.025 * fs), 8).astype(int)):
early[i] += amplitude * 0.85 ** k * (1.0 if k % 2 == 0 else -1.0)
cases = {
"single slope": tail,
"coupled volume": noise * np.sqrt(0.98 * energy(0.6) + 0.02 * energy(2.5)),
"strong early energy": early,
}
fig, axes = plt.subplots(1, 3, sharey=True, figsize=(13.0, 4.4))
for ax, (name, signal) in zip(axes, cases.items()):
res = room.room_parameters(signal, fs, limits=None)
room.decay_curve(signal, fs).plot(ax=ax)
ax.set_title(f"{name}\nEDT {res.edt[0]:.2f} T20 {res.t20[0]:.2f} "
f"T30 {res.t30[0]:.2f} C {res.curvature[0]:.0f} %")
plt.show()

How much decay range the noise floor allows

Section titled “How much decay range the noise floor allows”

A fit window is only as good as the decay range underneath it. The impulse-to-noise ratio (INR) is the level distance between the peak of the band-filtered IR and its noise floor; the fit window plus a safety margin must fit inside it, which is the ISO 3382 requirement of at least 35 dB of usable decay range for T20 and 45 dB for T30. An undersized range biases the fitted time upward, toward the flat tail the noise floor imposes on the decay curve, and the bias grows quietly before the fit visibly fails (Hak, Wenmaekers, & van Luxemburg, 2012). room_parameters reports that INR per band as dynamic_range and tightens the acceptance limits to 46 dB (T20) and 54 dB (T30) before flagging a value valid, so the residual truncation-and-compensation bias of a flagged-valid time stays inside 5 %: the relative JND Table A.1 tabulates for EDT and the value customarily applied to T20 and T30, for which the table gives none. When a band fails its flag, the order of remedies is: use T20 instead of T30 (its window needs 10 dB less range under the ISO minima, 8 dB under the tightened flags); raise the INR at acquisition, since doubling the sweep length or the number of synchronous averages buys 3 dB each time; and only then fall back to EDT, never to a fit stretched into the noise.

Where the noise floor comes from, and how to lose it. The background level is estimated from the last tenth of the onset-trimmed response, so the recording must actually contain the room’s steady background after the decay has finished (half a second or more of it). A file trimmed to the audible decay, or a length= set to the decay alone, makes the analyser measure decaying signal as if it were noise: the estimated floor sits too high, the truncation point moves earlier, the compensated tail is computed from the wrong crossing, and the per-band validity flags are applied to a fictitious dynamic range, all without an error or a warning. Record silence after the excitation (the same silence ISO 18233 6.2.2.3 asks for), keep it in the file, and let room_parameters find the crossing rather than trimming to taste. The diagnostic for the mistake: an implausibly large dynamic_range on a band that should be noisy, together with a curvature near zero, is the signature of a missing noise tail.

The whole budget fits in one drawing, and so does the bias it controls.

Level against time for one band-filtered impulse response: the peak at 0 dB, the squared impulse response falling as a straight line, a horizontal background-noise plateau at minus 55 decibels, a vertical bracket between the two labelled INR, the crossing of the fitted line with the noise estimate marked as the truncation point with the region beyond shaded as the compensated tail, and on the right a stacked ruler of the EDT, T20 and T30 evaluation windows each with a hatched 15 decibel margin extension and the 46 and 54 decibel flag thresholds marked against the INR bracketLevel against time for one band-filtered impulse response: the peak at 0 dB, the squared impulse response falling as a straight line, a horizontal background-noise plateau at minus 55 decibels, a vertical bracket between the two labelled INR, the crossing of the fitted line with the noise estimate marked as the truncation point with the region beyond shaded as the compensated tail, and on the right a stacked ruler of the EDT, T20 and T30 evaluation windows each with a hatched 15 decibel margin extension and the 46 and 54 decibel flag thresholds marked against the INR bracket

The level budget of one band. The INR is the whole vertical space you have; the evaluation window plus the 15 dB margin of ISO 3382-1 clause 5.3.3 has to fit inside it, which is where 35 dB (T20) and 45 dB (T30) come from. The integration stops where the fitted line crosses the background estimate, and the tail beyond is replaced by the exponential continuation of the fit, a compensation whose residual bias is positive, which is why phonometry adds a further 11 dB and 9 dB of headroom before it calls a band valid.

Percentage error of the fitted T20 and T30 against the returned dynamic range from 25 to 80 decibels for a synthetic one second decay: both errors are positive and grow as the dynamic range shrinks, T30 leaving the shaded plus or minus five per cent band below about 54 decibels and T20 below about 46 decibels, with vertical markers at the ISO minima of 35 and 45 decibels and the flagged-invalid region hatchedPercentage error of the fitted T20 and T30 against the returned dynamic range from 25 to 80 decibels for a synthetic one second decay: both errors are positive and grow as the dynamic range shrinks, T30 leaving the shaded plus or minus five per cent band below about 54 decibels and T20 below about 46 decibels, with vertical markers at the ISO minima of 35 and 45 decibels and the flagged-invalid region hatched

A synthetic = 1.0 s decay with its noise floor swept, measured with room_parameters. The bias is always positive: the flat noise tail can only make the fitted decay look longer, and it grows as the range shrinks: +0.4 % at 70 dB, +1.3 % (T30) at 61 dB, +3.1 % at 55 dB, +6.6 % at 46 dB and +8.8 % for T20 at 37 dB, just above the ISO minimum. That is the whole argument for the tightened flags: at ISO’s own 35 dB and 45 dB minima the surviving bias is already outside the 5 % JND, and only at 46 dB and 54 dB does it fall inside. Below about 34 dB the fit does not converge at all and the parameters return NaN, which is the visible failure the bias creeps up on.

Show the code for this figure
import matplotlib.pyplot as plt
# `np`, `fs` and `room` are the imports of the decay block above.
rng = np.random.default_rng(3382)
grid = np.arange(int(3.0 * fs)) / fs
clean = rng.standard_normal(grid.size) * np.exp(-3.0 * np.log(10.0) * grid)
noise = rng.standard_normal(grid.size)
peak = float(np.max(np.abs(clean)))
ranges, bias20, bias30 = [], [], []
for inr in range(25, 81, 3):
res = room.room_parameters(clean + noise * peak * 10 ** (-inr / 20.0),
fs, limits=None)
ranges.append(float(res.dynamic_range[0]))
bias20.append(100.0 * (float(res.t20[0]) - 1.0))
bias30.append(100.0 * (float(res.t30[0]) - 1.0))
fig, ax = plt.subplots()
ax.plot(ranges, bias20, "-o", label="T20")
ax.plot(ranges, bias30, "-s", label="T30")
ax.axhspan(-5.0, 5.0, alpha=0.15)
for limit, name in ((35.0, "ISO T20"), (45.0, "ISO T30"),
(46.0, "flag T20"), (54.0, "flag T30")):
ax.axvline(limit, ls=":", alpha=0.6)
ax.set_xlabel("dynamic_range (INR) [dB]")
ax.set_ylabel("Error of the fitted decay time [%]")
ax.legend()
plt.show()

Below the Schroeder frequency: modes, not decay

Section titled “Below the Schroeder frequency: modes, not decay”

Below the Schroeder frequency ( in m³, in s) these decay statistics stop telling the whole story: the field is ruled by discrete room modes. The constant is not a physical one. Each mode damped to a reverberation time has a half-power bandwidth of about hertz, the modal density grows as , and requiring roughly three modes inside one bandwidth gives : a choice of how much overlap counts as diffuse, which is why stricter texts quote 4000 for the same physics. The transition is soft, so a band straddling should be checked position by position rather than trusted, and the mode ladder of the image-sources guide enumerates what is actually down there. The simulation below drives the same rigid 5 m by 3.5 m room at its (2,1) mode and then between two modes; on resonance a standing-wave pattern with fixed nodal lines grows until it dominates the RMS pressure map, off resonance the room still responds, but the forced field stays weak and never organises into that (2,1) nodal pattern.

A 2D FDTD simulation of a rigid 5 by 3.5 metre room driven at the 84 Hz (2,1) mode and at an off-mode frequency side by side. On resonance a standing-wave pattern with fixed nodal lines grows until it dominates the RMS pressure map; off resonance the forced response stays weak and never organises into that nodal structure.

Download the animation (WebM)

A 2D FDTD simulation of a rigid 5 by 3.5 metre room driven at the 84 Hz (2,1) mode and at an off-mode frequency side by side. On resonance a standing-wave pattern with fixed nodal lines grows until it dominates the RMS pressure map; off resonance the forced response stays weak and never organises into that nodal structure.

Download the animation (WebM)

Parameter Type Units Range / default Notes
ir 1D array any non-silent Measured impulse response
fs int Hz > 0 Sample rate
limits (float, float) or None Hz default (125.0, 4000.0) Band-centre limits; None = broadband single band
fraction int — 1 (octave, default) / 3 (third) Bandwidth fraction
zero_phase bool — default False Forward-backward octave filtering (ISO 3382-2 §7.3 NOTE, which relaxes to ); removes the filter group delay before the backward integration and roughly halves the 125 Hz short-decay T30 bias (~+4.9 % → +2.4 % at = 0.2 s). decay_curve accepts it too

is the bandwidth-time product: the bandwidth of the analysis band in hertz and the reverberation time in that band, in seconds. The two symbols in that cell are not the same as the filter’s: ” = 0.2 s” means the reverberation time. The 125 Hz octave is about 88 Hz wide, so a 0.2 s decay gives and only just clears the ISO 3382-2 §7.3 threshold of 16; halve the decay or move to third octaves and it does not, which is the regime the NOTE’s relaxation exists for.

Returns a RoomAcousticsResult: frequencies (band centres, or None broadband), edt/t20/t30 (s), c50/c80 (dB), d50, ts (s), dynamic_range (dB, the per-band INR: band-filtered peak minus noise floor), the edt_valid/t20_valid/t30_valid flags (ISO 3382-1 §5.3.3: noise ≥ 25 dB below the peak for EDT, tightened to 46 dB for T20 and 54 dB for T30 so the tail-compensation bias of a flagged-valid value stays within the 5 % JND) and curvature % (values above 10 % flag a non-straight decay). decay_curve(ir, fs, band=None, fraction=1, zero_phase=False) returns just the decay curve for one band or the broadband response, a DecayCurve read by name as .times and .levels.

Everything above analyses one impulse response; a reported room parameter is never one impulse response. ISO 3382-2 clause 8 gives two admissible procedures, and the one this API supports is the first: the arithmetic mean of the individual reverberation times over all independent source-microphone combinations (the alternative, ensemble-averaging the decay curves themselves with their beginnings synchronised, is a different operation and is not what averaging t30 does). Four rules travel with it. The standard deviation across positions is part of the answer, not a diagnostic afterthought, and should be quoted whenever it exceeds the parameter’s JND. , the customary single-number room descriptor, is the mean of the 500 Hz and 1 kHz octaves. The energy parameters vary far more between seats than T30 does, so their spread is the interesting number rather than a nuisance. And a band that fails its validity flag at some positions must not be quietly averaged in: mask it, and say how many positions survived.

# `np`, `fs` and `room` are the imports of the decay block above.
# measured: one impulse response per source-microphone combination. Three
# synthetic decays of 0.95, 1.00 and 1.08 s stand in for three positions.
rng = np.random.default_rng(3382)
grid = np.arange(int(3.0 * fs)) / fs
responses = [rng.standard_normal(grid.size) * np.exp(-3.0 * np.log(10.0) / t60 * grid)
for t60 in (0.95, 1.00, 1.08)]
results = [room.room_parameters(each, fs) for each in responses]
t30 = np.array([res.t30 for res in results])
valid = np.array([res.t30_valid for res in results])
mean = np.nanmean(np.where(valid, t30, np.nan), axis=0)
spread = np.nanstd(np.where(valid, t30, np.nan), axis=0)
print(np.round(mean, 2)) # [0.98 1. 1.02 1.02 1.03 1.01] s
print(np.round(100.0 * spread / mean, 1)) # [4.5 4.3 4.8 5.1 4.9 4.8] %
print(int(valid.sum(axis=0).min())) # 3 positions survived the flag

Here the spread runs 4 to 5 %, right at the JND, so it belongs in the report beside the mean: three positions that agree to 5 % are three positions that disagree audibly. Note also what the mask is for: np.nanmean over the masked array is the arithmetic mean clause 8 a) prescribes, but only over the combinations whose band actually passed, and the surviving count is part of the result.

A measured chapel against its published parameters

Section titled “A measured chapel against its published parameters”

Everything above runs on synthetic decays. The repository also keeps a real one: the impulse response of the St. Nicholas Chapel of Aachen Cathedral, a coupled Gothic chapel of about 3050 m³ that FH Aachen measured to ISO 3382-1 with a dodecahedral source and an exponential sweep, source 1 and receiver 1 standing 2.56 m apart. The deposit publishes the room parameters its authors computed with their own software, per one-third-octave band, so the same response can be run through room_parameters and set against them. The yardstick is the one ISO 3382-1 gives for whether two values of a parameter differ audibly: the just noticeable differences (JND) of Table A.1, 5 % of the value for EDT, 1 dB for C80, 0.05 for D50 and 10 ms for the centre time Ts, which room.TABLE_A1 holds.

The data. Omnidirectional channel of the S1-Do_R1 impulse response from M. Zerwas and S. Kayku, ‘Room acoustic measurement and decay analysis data of the St. Nicholas Chapel, Aachen Cathedral’, Zenodo, 2026, DOI 10.5281/zenodo.20428705, licensed CC BY 4.0. Change: one channel of four extracted and repackaged.

The file is read from a checkout of the repository, where the oracle test reads it too; a wheel does not install the test data.

import pathlib
import numpy as np
import phonometry
from phonometry import room
# The committed response lives in the repository, so this runs from a checkout
# (or an editable install); with your own measured response, load that instead.
repo = pathlib.Path(phonometry.__file__).resolve().parents[2]
with np.load(repo / "tests/data/audio/nikolauskapelle/s1_do_r1_w.npz") as npz:
chapel_ir, chapel_fs = npz["w"].astype(np.float64), int(npz["fs"])
chapel = room.room_parameters(chapel_ir, chapel_fs, limits=(100.0, 5000.0), fraction=3)
bands = [100, 125, 160, 200, 250, 315, 400, 500, 630, 800,
1000, 1250, 1600, 2000, 2500, 3150, 4000, 5000]
# The deposit's workbook, row S1-Do_R1, one value per band from 100 Hz to 5 kHz.
published = {
"EDT": [3.3121, 3.4835, 4.6281, 4.2371, 4.3389, 3.8817, 3.6988, 3.8547, 4.4786,
4.9678, 3.8599, 3.4878, 2.5415, 2.6911, 2.2371, 1.7814, 1.4736, 1.198],
"T20": [4.3325, 4.4211, 4.2996, 4.7631, 4.944, 4.9925, 4.7133, 4.7381, 4.649,
4.5321, 4.533, 4.1439, 3.8006, 3.3868, 2.8327, 2.4767, 1.9659, 1.5017],
"T30": [4.7058, 4.7348, 4.4475, 4.6511, 4.6734, 4.8526, 4.7217, 4.8584, 4.7744,
4.6152, 4.4409, 4.1049, 3.8218, 3.3945, 2.9959, 2.4781, 2.0283, 1.5823],
"C50": [-0.8857, -1.6775, -1.1019, -0.8407, -2.4858, -1.0162, 0.0553, -0.7639,
-2.097, 0.0628, -1.4394, 0.3681, 1.4292, 1.7993, 1.1941, 1.1194, 2.5834, 4.6042],
"C80": [-0.36, -0.8159, -1.2315, -0.1926, -1.8477, 0.7525, 0.9272, -0.171, -1.5149,
0.9836, 0.023, 1.7805, 2.8353, 3.2893, 3.2199, 2.5758, 4.344, 6.177],
"D50": [0.4552, 0.4075, 0.4352, 0.4514, 0.3617, 0.4439, 0.5076, 0.4519, 0.3836,
0.5026, 0.4179, 0.5206, 0.5817, 0.602, 0.5684, 0.564, 0.6443, 0.743],
# Ts in milliseconds, as the workbook gives it, turned into seconds.
"Ts": [ms / 1000 for ms in (226.79, 220.39, 232.98, 189.96, 260.19, 182.72, 162.69, 202.2,
242.02, 213.45, 195.51, 153.0, 114.73, 104.2, 100.94, 90.64,
64.43, 44.36)],
}
computed = {"EDT": chapel.edt, "T20": chapel.t20, "T30": chapel.t30,
"C50": chapel.c50, "C80": chapel.c80, "D50": chapel.d50, "Ts": chapel.ts}
print(f"T30 at 1 kHz: {chapel.t30[10]:.2f} s, published {published['T30'][10]:.2f} s")
# T30 at 1 kHz: 4.46 s, published 4.44 s
# ISO 3382-1 Table A.1 gives the just noticeable difference (JND) of EDT, C80,
# D50 and Ts; T20, T30 and C50 have no row of their own, and the test holds
# them to the JND of EDT (5 %) and of C80 (1 dB).
jnd_row = {"EDT": "EDT", "T20": "EDT", "T30": "EDT",
"C50": "C80", "C80": "C80", "D50": "D50", "Ts": "Ts"}
for name, values in published.items():
row = room.TABLE_A1[jnd_row[name]]
gap = computed[name] - np.asarray(values)
if row.relative_jnd:
gap = gap / np.asarray(values)
in_jnds = gap / row.just_noticeable_difference
worst = int(np.argmax(np.abs(in_jnds)))
print(f"{np.sum(np.abs(in_jnds) <= 1)} of {len(bands)} bands within one JND for {name}, "
f"worst {in_jnds[worst]:+.1f} JND at {bands[worst]} Hz")
# 17 of 18 bands within one JND for EDT, worst +1.3 JND at 5000 Hz
# 18 of 18 bands within one JND for T20, worst +0.5 JND at 250 Hz
# 18 of 18 bands within one JND for T30, worst +0.6 JND at 250 Hz
# 17 of 18 bands within one JND for C50, worst -1.5 JND at 100 Hz
# 18 of 18 bands within one JND for C80, worst -0.9 JND at 100 Hz
# 17 of 18 bands within one JND for D50, worst -1.8 JND at 100 Hz
# 17 of 18 bands within one JND for Ts, worst +1.3 JND at 250 Hz

T20, T30 and C80 land on the published value within one just noticeable difference in all 18 bands, EDT, C50, D50 and Ts in 17 of them, and T30 at 1 kHz reads 4.46 s against the published 4.44 s. The four misses are the two edge cases the test documents and holds at three JNDs instead of one. EDT at 5 kHz is +1.3 JND: air absorption shortens the decay at the top of the range, so the 0 to −10 dB fit has very little of it to work with. And the quantities that hang on where the response starts: the early-energy ratios at 100 Hz, C50 at −1.5 JND and D50 at −1.8, with C80 just inside at −0.9, and the centre time Ts at 250 Hz, +1.3 JND. At 100 Hz a one-third-octave filter’s group delay is comparable to the 50 ms and 80 ms early/late split, so where each implementation puts the start of the response decides most of the ratio; Ts is a time counted from that same start, so it moves with it. Two independent implementations agreeing that closely on decays this long is as strong a statement as the parameters’ definitions allow.

ISO 3382 per-band parameters of the St. Nicholas Chapel in one-third octaves from 100 Hz to 5 kHz: EDT, T20 and T30 bars falling from about 4.7 s to 1.5 s, each topped by an open marker at the value the deposit publishes, over a second panel where the computed C50 and C80 curves follow the published values within a decibel in every band but one, C50 at 100 HzISO 3382 per-band parameters of the St. Nicholas Chapel in one-third octaves from 100 Hz to 5 kHz: EDT, T20 and T30 bars falling from about 4.7 s to 1.5 s, each topped by an open marker at the value the deposit publishes, over a second panel where the computed C50 and C80 curves follow the published values within a decibel in every band but one, C50 at 100 Hz

The chapel through the result’s own .plot(), with the deposit’s published row drawn over it as open markers. The bars and the curves are what room_parameters computes from the impulse response alone; the markers are the authors’ numbers, within one just noticeable difference of the bars and the lines in every band but the ones the prints above name.

Show the code for this figure
import matplotlib.pyplot as plt
import numpy as np
# `chapel`, `published` and `bands` come from the snippet above.
# One line: the result draws its own bars and curves; the published row goes on top.
ax_times, ax_clarity = chapel.plot()
x = np.arange(len(bands))
for offset, name in [(-0.27, "EDT"), (0.0, "T20"), (0.27, "T30")]:
ax_times.plot(x + offset, published[name], "o", mfc="none", color="k",
label="Published" if name == "EDT" else None)
for name in ["C50", "C80"]:
ax_clarity.plot(x, published[name], "o", mfc="none", label=f"{name} published")
ax_times.legend()
ax_clarity.legend()
plt.show()
# By hand: the decay times as bars and the clarities as lines, the published
# row as open markers over them, from the result's fields.
fig, (top, bottom) = plt.subplots(2, 1, sharex=True)
for offset, name, values in [(-0.27, "EDT", chapel.edt), (0.0, "T20", chapel.t20),
(0.27, "T30", chapel.t30)]:
top.bar(x + offset, values, width=0.27, label=name)
top.plot(x + offset, published[name], "o", mfc="none", color="k")
for name, values in [("C50", chapel.c50), ("C80", chapel.c80)]:
line, = bottom.plot(x, values, "o-", label=name)
bottom.plot(x, published[name], "o", mfc="none", color=line.get_color())
bottom.set_xticks(x, [str(band) for band in bands], rotation=45)
bottom.set_xlabel("Frequency [Hz]")
top.set_ylabel("Decay time [s]")
bottom.set_ylabel("Clarity [dB]")
top.legend()
bottom.legend()
plt.show()

RoomAcousticsResult.report(path) renders a one-page PDF fiche laid out like a room-acoustics measurement report (a performance space per ISO 3382-1:2009 or an ordinary room per ISO 3382-2:2008, both evaluated by the integrated impulse-response method): a standard-basis line, an optional metadata header block, the full-width per-band parameter table (, , EDT, , , , ) above the result’s own per-band decay-time plot (.plot()), the boxed mid-frequency reverberation time (the mean of the 500 Hz and 1000 Hz octave , the customary room descriptor) with the mid-frequency EDT alongside, and a footer with the fixed disclaimer. ISO 3382-1/-2 are characterisation standards with no intrinsic pass/fail, so a verdict row appears only when a target reverberation time is supplied through the metadata’s requirement field (ReportMetadata(requirement=...), read as the maximum acceptable value of whichever descriptor the box carries, the usual form of a classroom or open-plan limit). A broadband result has no 500 Hz / 1000 Hz octaves to average, so its box and verdict fall back to the plain broadband with no “500-1000 Hz” claim, and so do a band range that misses those two octaves and one that reaches them with a non-evaluable in either, which fall back to the first band whose could be evaluated. It uses the same ReportMetadata container and rendering engine as the ISO 11654 absorption fiche; the room-specific fields room_volume, n_source_positions and n_receiver_positions populate the header (ISO 3382 requires the room volume and the number of source and microphone positions to be reported), alongside test_room (the room name), specimen (the room description and occupancy state), area, instrumentation, temperature_c, relative_humidity_percent, static_pressure_kpa, measurement_standard, test_date, laboratory, operator, report_id and notes. Passing metadata=None produces a bare characterisation fiche. Rendering needs reportlab and, for the figure the fiche embeds, matplotlib (pip install "phonometry[report,plot]"); only engine="reportlab" is supported. The fiche renders in English by default; pass language="es" for a Spanish fiche (translated fixed strings and a comma decimal separator).

from phonometry import room, ReportMetadata
result = room.room_parameters(ir, fs) # octave bands 125 Hz - 4 kHz
result.report(
"room_fiche.pdf",
metadata=ReportMetadata(
# The room condition clause 9.2 e) asks for: furnishing state and the
# number of persons present (clause 4.1 allows at most two to stand in
# for the unoccupied state).
specimen="Small auditorium, fully furnished, unoccupied (1 operator present)",
test_room="Auditorium A",
room_volume=2830.0, area=340.0,
n_source_positions=2, n_receiver_positions=8,
# Clause 9.2 i) requires the accuracy grade to be stated.
measurement_standard="ISO 3382-2 (engineering grade)",
temperature_c=21.0, relative_humidity_percent=45.0,
laboratory="Phonometry Reference Laboratory",
notes="16 source-microphone combinations, 2 decays each; "
"positions and heights on the attached scaled plan",
requirement=1.3, # adds a verdict against a target T_mid
),
) # T_mid + the per-band parameter table

Three of those entries are there because ISO 3382-2 clause 9.2 asks for them and the fiche has no dedicated field for any of them. Clause 9.2 i) requires the degree of precision (survey, engineering or precision), so it belongs in measurement_standard rather than being left implied. Clause 9.2 e) requires the condition of the room, furnishing and the number of persons present, and clause 4.1 fixes what “unoccupied” may mean: no people, or at most two allowed to stand in for the unoccupied state unless otherwise specified; that goes in specimen. And clause 9.2 c) and i) together require a scaled sketch plan carrying the source and microphone positions with their heights, which the fiche does not draw: produce it alongside and reference it, as notes does here. Clause 9.2 f) adds the temperature and relative humidity for the precision method only, which the climate fields already carry.

The example fiche is regenerated with make reports and kept rendered in the repository; click the preview to open the PDF.

ISO 3382 room acoustic parameters example report (PDF)

One-page room-acoustics fiche: a metadata header (client, room, description, room volume, floor area, source and microphone positions, instrumentation, temperature, humidity and pressure), the octave-band parameter table (T20, T30, EDT, C50, C80, D50 and Ts from 125 Hz to 4 kHz) above the per-band decay-time bar plot, the boxed mid-frequency reverberation time T_mid = 1.15 s with the mid-frequency EDT alongside, and a PASS verdict against the 1.3 s target.

Download the report (PDF)

Room acoustic parameters fiche (RoomAcousticsResult.report), T_mid and the per-band table.

With the parameters in hand, the two neighbouring links in the chain are the acquisition that produced the impulse response (Measuring the Room Impulse Response) and, for an open-plan floor rather than a closed room, the ISO 3382-3 line of workstations (Open-Plan Office Acoustics).

  • Covered

    ISO 3382-1:2009 (clause 5.3.3 backward integration and the Annex A parameter formulae) and ISO 3382-2:2008 (Clause 6 and the 5.3.3 dynamic-range criterion), computed by room.room_parameters/room.decay_curve: EDT, T20, T30, C50, C80, D50, Ts and the per-band validity and curvature flags, with the optional zero-phase band filtering of the ISO 3382-2 §7.3 NOTE, and the one-page ISO 3382 fiche through .report().

  • Not covered

    ISO 3382-1 Annex A also defines the sound-strength and the binaural parameters LF/LFC and IACC; phonometry implements only the reverberance and clarity family named above (EDT, T20, T30, C50, C80, D50, Ts). The acquisition of the impulse response (ISO 18233) and the open-plan quantities (ISO 3382-3) have guides of their own, Measuring the Room Impulse Response and Open-Plan Office Acoustics. The ISO 354 reverberation-room absorption that consumes the measured / pair lives in Sound Absorption Measurement and Rating.

Each band of the impulse response is turned into a decay curve by Schroeder backward integration of the squared IR, and a least-squares line fitted over an evaluation range is extrapolated to a full 60 dB drop, (ISO 3382-1/2): EDT over 0 to −10 dB (perceived reverberance), T20 over −5 to −25 dB and T30 over −5 to −35 dB.

How much decay range do I need for a valid T20 or T30?

Section titled “How much decay range do I need for a valid T20 or T30?”

The fit window plus a safety margin must fit inside the impulse-to-noise ratio (res.dynamic_range), the level distance between the band-filtered IR peak and its noise floor: ISO 3382 requires at least 35 dB of usable decay range for T20 and 45 dB for T30. An undersized range biases the fitted time upward, so room_parameters tightens its validity flags to 46 dB and 54 dB, keeping the bias inside the 5 % JND. The noise floor is estimated from the last tenth of the response, so the recording must contain silence after the decay.