# Reverberation-time prediction (Sabine, Eyring, Arau)

> One page of the phonometry documentation, as Markdown. Index: https://docs.jmrp.io/phonometry/llms.txt

Canonical: https://docs.jmrp.io/phonometry/buildings/rooms/reverberation-prediction/
Language: en
Alternate: https://docs.jmrp.io/phonometry/es/buildings/rooms/reverberation-prediction/index.md
Version: phonometry 4.0.0
Install: pip install "phonometry[full]"
Author: José Manuel Requena Plens (https://orcid.org/0000-0003-1250-6212)
Published: 2026-07-11
Updated: 2026-10-28T03:12:25Z
License: MIT (https://github.com/jmrplens/phonometry/blob/main/LICENSE)
Summary: Reverberation-time prediction from volume, areas and absorption: Sabine, Eyring, Millington-Sette, Fitzroy and Arau-Puchades, with air absorption.

---

The **reverberation time** $T$, the time for the sound-energy level to fall by
60 dB after the source stops, is predicted here from a room's **volume**,
**boundary areas** and the **sound-absorption coefficients** of its surfaces,
through the classical statistical-acoustics formulae. This is the design-stage
counterpart of the *measured* reverberation time of
[Room acoustics](https://docs.jmrp.io/phonometry/buildings/rooms/room-acoustics/) (ISO 3382) and complements
the EN 12354-6 model of
[Sound absorption in enclosed spaces](https://docs.jmrp.io/phonometry/buildings/rooms/enclosed-space-absorption/),
which specialises the same physics to that standard's Clause 4.

phonometry offers five models, ordered by how much they account for a
**non-uniform** absorption distribution:

| Model | How the absorption enters | Best for |
|:---|:---|:---|
| **Sabine** | absorption term $A = \sum_i S_i\alpha_i$ | low, uniform absorption |
| **Eyring** (Norris-Eyring) | absorption term $-S\ln(1-\bar\alpha)$ | strong, uniform absorption |
| **Millington-Sette** | absorption term $-\sum_i S_i\ln(1-\alpha_i)$ | a few very absorptive surfaces |
| **Fitzroy** | area-weighted **arithmetic** mean of three axial Eyring times | anisotropic rooms |
| **Arau-Puchades** | area-weighted **geometric** mean of the same three | anisotropic rooms (author-preferred) |

The first three models differ only in the denominator of
$T = k\,V / (\text{term} + 4mV)$, with the Sabine constant
$k = 24\ln 10 / c_0$ (so $k = 0.161$ for $c_0 = 343\ \mathrm{m/s}$) and the
air-absorption term $4mV$. Fitzroy and Arau-Puchades work differently: they
compute three axial Eyring *times* and combine those, so their entries are
reverberation times and cannot be substituted into that denominator (§2).

EN 12354-6 pins the same constant differently. It rounds $24\ln 10$ to 55.3 and
fixes $c_0 = 345.6\ \mathrm{m/s}$ so that the factor is exactly the traditional
0.16, which makes the
[enclosed-space model](https://docs.jmrp.io/phonometry/buildings/rooms/enclosed-space-absorption/)
return reverberation times about 0.7 % shorter than
`sabine_reverberation_time` for the same room (0.608 s against 0.612 s) for the
8 × 5 × 3 m shoebox of §1. It is a rounding convention, not a physical
disagreement: passing `speed_of_sound=345.6` to the classical function brings
the two within 0.07 %, the residue of rounding $24\ln 10 = 55.262$ to 55.3.

![Reverberation time per octave band for a room with an absorptive floor and ceiling but hard walls, computed by five models](https://docs.jmrp.io/phonometry/media/reverberation_models.svg)

*A 10 × 7 × 3.5 m room with a carpeted floor and an acoustic ceiling between
hard end walls, run through all five models. Fitzroy sits well above the rest
because a single wall pair carries almost all the absorption; Sabine and Eyring
bracket the middle, and Arau-Puchades tempers Fitzroy's over-prediction with a
geometric mean. Where the five spread like this, the room is telling you its
field is not diffuse (§4): the spread is the diagnostic, not a defect of one
formula.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt
from phonometry import environment, room

# A 10 x 7 x 3.5 m room: hard end walls, lightly treated side walls and a
# very absorptive floor/ceiling pair (carpet plus an acoustic ceiling).
bands = [125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0]
alpha_x = [0.06, 0.07, 0.08, 0.09, 0.10, 0.10]
alpha_y = [0.12, 0.14, 0.16, 0.18, 0.20, 0.20]
alpha_z = [0.30, 0.50, 0.65, 0.78, 0.82, 0.80]
m = environment.air_power_attenuation_per_m(bands, temperature_c=20.0, relative_humidity_percent=50.0)   # air at 20 C / 50 % RH
res = room.reverberation_time_models((10.0, 7.0, 3.5),
                                     (alpha_x, alpha_y, alpha_z),
                                     air_attenuation_per_m=m, frequencies=bands)
res.plot()   # the five model curves per band
plt.show()
```

## 0. Assembling the inputs

All five models consume two things a drawing does not hand you directly: a list
of areas with a coefficient each, and (for Fitzroy and Arau-Puchades) one mean
coefficient per pair of opposing walls. Almost every wrong prediction is a wrong
take-off rather than a wrong formula, so this step deserves as much care as the
choice of model.

**Areas.** Use the *internal* boundary areas of the room, measured to the
finished surfaces, and give every material its own row: a wall carrying a window
and a door becomes three rows whose areas sum to the wall. EN 12354-6 Annex E
case 3 does exactly this, splitting a 10.90 m² long wall into 9.81 m² lined at
$\alpha_\mathrm{s}$ = 0.85 plus 1.09 m² bare at 0.04. Do not average a lining into its
wall by hand; the formulae already do the area weighting, and Millington-Sette
does it non-linearly, so pre-averaging changes the answer.

**Coefficients.** These are random-incidence values measured to ISO 354 on a
10-12 m² sample in a stated mounting, not material constants
([Sound Absorption Measurement and Rating](https://docs.jmrp.io/phonometry/materials/absorbers/absorption-measurement/)).
Mounting depth and edge condition change them substantially, so a coefficient
quoted for a product bonded direct to the substrate does not describe the same
product on a 200 mm plenum. ISO 354 delivers one-third-octave data and an
octave-band run takes the arithmetic mean of the three thirds. A datasheet's
single-number rating ($\alpha_\mathrm{w}$, NRC, SAA) is *not* a substitute for the
per-band table (see the
[EN 12354-6 guide](https://docs.jmrp.io/phonometry/buildings/rooms/enclosed-space-absorption/#3-where-the-input-data-comes-from),
which states the prohibition normatively). Audience and seating are rated per
square metre of the floor area they occupy, not per person. Feeding
reverberation-room data back into Sabine is self-consistent because the ISO 354
coefficient is *defined* through Sabine's formula, which is also why a
coefficient above 1.0 is a documented outcome rather than an error (§4).

**Wall-pair means.** `fitzroy_reverberation_time` and
`arau_puchades_reverberation_time` take one mean coefficient per *opposing
pair*, in the order $(\bar\alpha_x, \bar\alpha_y, \bar\alpha_z)$ where
$\bar\alpha_i$ belongs to the pair perpendicular to axis $i$. Each is the
area-weighted mean over both surfaces of the pair. In a shoebox the two
surfaces have equal areas, so it reduces to their arithmetic mean, but not once
one of them is split by a window or a lining. Absorption that is patchy *within*
one pair (an absorptive wall facing a hard one) is outside both models even
though the code still returns a number; that case belongs to §4's list of
failures.

```python
# `room` is the import of the figure block above.
# A 9 x 7 x 3 m classroom (V = 189 m3), 1 kHz octave band, taken off by material.
surfaces = [
    (63.0, 0.05),   # floor, vinyl on concrete
    (63.0, 0.70),   # ceiling, 20 mm acoustic tile
    (12.0, 0.04),   # long wall A: glazing
    (15.0, 0.05),   # long wall A: plaster around the glazing
    (2.0, 0.10),    # long wall B: wooden door
    (25.0, 0.05),   # long wall B: plaster
    (21.0, 0.05),   # short wall, plaster
    (21.0, 0.05),   # short wall, plaster
]
print(round(room.sabine_reverberation_time(189.0, surfaces), 2))   # 0.59 s

def pair_mean(rows):                       # area-weighted mean over one pair
    area = sum(s for s, _ in rows)
    return sum(s * a for s, a in rows) / area

means = (pair_mean(surfaces[6:8]),         # x-pair: the two short walls
         pair_mean(surfaces[2:6]),         # y-pair: the two long walls
         pair_mean(surfaces[0:2]))         # z-pair: floor and ceiling
print([round(a, 3) for a in means])                                # [0.05, 0.05, 0.375]
print(round(room.fitzroy_reverberation_time((9.0, 7.0, 3.0), means), 2))       # 1.33 s
print(round(room.arau_puchades_reverberation_time((9.0, 7.0, 3.0), means), 2)) # 0.76 s
```

One take-off, both input forms. The spread it produces (0.59 s from Sabine
against 1.33 s from Fitzroy) is not a bug: all of this room's absorption sits
on one wall pair, which is exactly the case §2 exists for and §4 warns about.

## 1. Sabine, Eyring and Millington-Sette

Every model in the family follows from one picture. In a diffuse field a ray
travels a **mean free path** $4V/S$ between reflections, so it meets a boundary
$cS/4V$ times per second and gives up part of its energy each time. Take that
loss as $\alpha$ per reflection and linearise it and you get **Sabine**: the
decay rate is proportional to $\sum_i S_i\alpha_i$, and 60 dB of decay takes
$k V / A$ seconds. Keep the loss multiplicative instead: after $n$ reflections
the energy is down by $(1-\bar\alpha)^n$, and the logarithm of that factor
gives **Eyring**'s $-S\ln(1-\bar\alpha)$, which is why Eyring is always the
shorter of the two and why the two converge as $\bar\alpha \to 0$. Do the same
bookkeeping surface by surface rather than on the mean and you get
**Millington-Sette**, which collapses to zero the moment one surface absorbs
perfectly, because a ray that strikes it never returns. The air term $4mV$ is
the same path length applied to the medium instead of to the boundary.

The three statistical models take the room volume and a list of
`(area, absorption_coefficient)` surfaces. **Sabine** is exact only for low,
uniform absorption; **Eyring** replaces the absorption area by
$-S\ln(1-\bar\alpha)$ and is correct where Sabine overestimates $T$;
**Millington-Sette** sums the Eyring term surface by surface, so a single
perfectly absorbing surface drives $T$ to zero.

$$
T_{\text{Sab}} = \frac{k V}{\sum_i S_i\alpha_i}, \qquad
T_{\text{Eyr}} = \frac{k V}{-S\ln(1-\bar\alpha)}, \qquad
T_{\text{Mil}} = \frac{k V}{-\sum_i S_i\ln(1-\alpha_i)}.
$$

```python
from phonometry import room

# A shoebox 8 x 5 x 3 m (V = 120 m3, S = 158 m2), uniform alpha = 0.2.
surfaces = [(40.0, 0.2), (40.0, 0.2), (24.0, 0.2),
            (24.0, 0.2), (15.0, 0.2), (15.0, 0.2)]
print(round(room.sabine_reverberation_time(120.0, surfaces), 3))            # 0.612 s
print(round(room.eyring_reverberation_time(120.0, surfaces), 3))            # 0.548 s
print(round(room.millington_sette_reverberation_time(120.0, surfaces), 3))  # 0.548 s
```

For a **uniform** distribution Eyring and Millington-Sette coincide, and both
fall below Sabine; Sabine's over-estimate at high absorption is the reason
Eyring exists. As $\alpha \to 0$, Eyring reduces to Sabine. Air absorption
enters every model through the power (intensity) attenuation coefficient $m$,
in reciprocal metres, which ISO 354 8.1.2.1 writes as
$m = \alpha / (10 \lg \mathrm{e})$ from the ISO 9613-1 $\alpha$ in dB/m:

```python
from phonometry import environment, room

m = environment.air_power_attenuation_per_m(2000.0, temperature_c=20.0, relative_humidity_percent=50.0)
surfaces = [(40.0, 0.3), (40.0, 0.3), (24.0, 0.3),
            (24.0, 0.3), (15.0, 0.3), (15.0, 0.3)]
print(round(room.eyring_reverberation_time(120.0, surfaces, air_attenuation_per_m=m), 3))
# 0.337 s, against 0.343 s with the air term omitted
```

Six milliseconds on a third of a second is why the air term is usually ignored
in a room this size, and it is a scaling argument rather than a rule of thumb:
$4mV$ grows with the **volume** while the boundary term grows with the
**area**, so the air's share rises with the room's linear dimension. In the same
120 m³ room the 2 kHz correction is 1.9 %; in a 20 000 m³ concert hall with
$\bar\alpha = 0.15$ the 4 kHz Eyring time falls from 4.31 s to 2.49 s at
20 °C / 50 % RH (42 %) and to 1.52 s if the hall is dry at 20 % RH. That is
the second half of the behaviour: $m$ rises steeply with frequency and falls
with humidity, so the driest condition is the worst case and a hall that
measures well in a humid summer can sound noticeably brighter in winter. The
coefficient itself comes from the ISO 9613-1
[atmospheric absorption](https://docs.jmrp.io/phonometry/environment/propagation/outdoor-propagation/)
model, evaluated at the room's own temperature and humidity.

Every statistical model also assumes a **diffuse field**, and low
frequencies break that assumption first: below the **Schroeder frequency**
$f_\mathrm{s} \approx 2000\sqrt{T/V}$ ($V$ in m³, $T$ in s: 141 Hz for a 200 m³
classroom with $T$ = 1 s) the room responds as a set of discrete modes, not as
a reverberant mixture. `room.schroeder_frequency(T, V)` computes it, and
[the modes of a rectangular room](https://docs.jmrp.io/phonometry/buildings/rooms/room-image-sources/#3-modes-of-a-rectangular-room)
enumerate what is down there. The
2D FDTD simulation below drives a rigid 5 m by 3.5 m room exactly on its
(2,1) mode and then between two modes; the standing-wave pattern that
builds up on resonance is what Sabine and Eyring cannot see.

**Video:** [On-mode versus off-mode drive in a rigid room (2D FDTD)](https://docs.jmrp.io/phonometry/media/anim_fdtd_room_modes.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.

## 2. Fitzroy and Arau-Puchades (anisotropic rooms)

When the absorption is concentrated on one axis (a carpeted floor and an
acoustic ceiling against otherwise hard walls), a single mean $\bar\alpha$
misrepresents the field. **Fitzroy** and **Arau-Puchades** split a rectangular
(shoebox) room into the three pairs of opposing walls and combine the *axial*
Eyring reverberation times $T_i$ (each using the whole surface $S$ and the mean
absorption $\bar\alpha_i$ of the wall pair perpendicular to axis $i$):

$$
T_{\text{Fitz}} = \sum_i \frac{S_i}{S}\,T_i \quad(\text{arithmetic}), \qquad
T_{\text{Arau}} = \prod_i T_i^{\,S_i/S} \quad(\text{geometric}).
$$

Each $T_i$ has a physical reading: it is the reverberation time the room *would*
have if every one of its boundaries absorbed like the wall pair on axis $i$, so
it is the decay experienced by sound bouncing predominantly along that axis,
which is why the formula uses the whole surface $S$ and not $S_i$. The weight
$S_i/S$ is then the fraction of the boundary that steers energy into that axis.
A shoebox with a soft floor and ceiling between hard walls therefore has three
coexisting decay rates, and the two models differ only in how they blend them:
Fitzroy averages them arithmetically, so the slowest axis dominates and the
result runs high when one pair is very reflective; Arau-Puchades averages them
geometrically, which lets the fastest axis pull the answer down. Both assume a
rectangular room and absorption uniform *within* each pair, so a room with one
absorptive wall facing a hard one of the same pair is outside the model even
though the code still returns a number (§0).

```python
from phonometry import room

# 8 x 5 x 3 m room, absorptive x-wall pair (alpha 0.5), hard elsewhere (0.1).
dims = (8.0, 5.0, 3.0)
absorption = (0.5, 0.1, 0.1)   # mean alpha of the (x, y, z) wall pairs
print(round(room.arau_puchades_reverberation_time(dims, absorption), 3))  # 0.812 s
print(round(room.fitzroy_reverberation_time(dims, absorption), 3))        # 0.974 s
```

By the arithmetic-geometric-mean inequality the Arau-Puchades time never exceeds
the Fitzroy time; Fitzroy is known to over-predict when one wall pair is very
reflective, which is why Arau-Puchades recommends the geometric mean. Both
reduce exactly to Eyring for a uniform absorption distribution.

The misrepresentation these two models exist for can be run as an
experiment, and the clip below does. One flat 8 × 2.5 m section is
simulated twice with the same **total** statistical absorption: spread over
all four edges, or concentrated on the floor-and-ceiling pair (the carpeted
floor and acoustic ceiling of this page's own example, α ≈ 0.79) between
hard ends. Any formula built on a single mean $\bar\alpha$ sees the same
number in both rooms, so the Sabine and Eyring predictions (evaluated for
the two-dimensional section, whose mean free path is $\pi S/P$) form one
band shared by both. The spread room decays inside it, a straight line at
T = 160 ms against Sabine's 201. The concentrated room's decay has no
single slope, 183 ms early and 236 ms late, and it finishes 10 dB above its
twin; the surviving field is visible in the RMS panel as a standing pattern
between the hard ends, nearly uniform from floor to ceiling, where sound runs
parallel to the absorber, brushing it at grazing incidence instead of
striking it, which is the field §4 describes in disproportionate rooms.

**Video:** [Same total absorption, two decays: placement in a flat room (2D FDTD)](https://docs.jmrp.io/phonometry/media/anim_fdtd_absorption_placement.webm): A 2D FDTD simulation of a flat 8 by 2.5 metre section run twice with the same total statistical absorption, spread over all four edges or concentrated on the floor and ceiling between hard ends. A lower axis races the two measured energy decays through the shared Sabine-Eyring band: the spread room decays inside the band at 160 ms, the concentrated room shows two slopes of 183 and 236 ms and finishes 10 dB louder, and its RMS map keeps a striped grazing field running parallel to the absorbing pair.

## 3. Comparing the five models per band

`reverberation_time_models` builds the six boundary surfaces of a rectangular
room from its dimensions and the three wall-pair mean absorptions, then
evaluates all five models on a common footing and returns a
`ReverberationModelResult` whose `.plot()` draws the figure above.

```python
from phonometry import room

# 10 x 7 x 3.5 m room, absorptive floor/ceiling against harder walls.
res = room.reverberation_time_models(
    (10.0, 7.0, 3.5),
    (
        [0.06, 0.07, 0.08, 0.09, 0.10, 0.10],   # x-pair: hard end walls
        [0.12, 0.14, 0.16, 0.18, 0.20, 0.20],   # y-pair: lightly treated walls
        [0.30, 0.50, 0.65, 0.78, 0.82, 0.80],   # z-pair: carpet + acoustic ceiling
    ),
    frequencies=[125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0],
)
print(res.sabine.round(2))         # [0.74 0.47 0.37 0.31 0.3  0.3 ]
print(res.arau_puchades.round(2))  # [0.79 0.51 0.38 0.29 0.26 0.27]
print(res.fitzroy.round(2))        # [1.02 0.79 0.66 0.57 0.51 0.51]
res.plot()   # the five model curves per band (the figure above)
```

**Show the code for this figure**

```python
import matplotlib.pyplot as plt
from phonometry import environment, room

m = environment.air_power_attenuation_per_m(
    [125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0],
    temperature_c=20.0,
    relative_humidity_percent=50.0,
)
room.reverberation_time_models(
    (10.0, 7.0, 3.5),
    (
        [0.06, 0.07, 0.08, 0.09, 0.10, 0.10],
        [0.12, 0.14, 0.16, 0.18, 0.20, 0.20],
        [0.30, 0.50, 0.65, 0.78, 0.82, 0.80],
    ),
    air_attenuation_per_m=m,
    frequencies=[125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0],
).plot()
plt.show()
```

Sabine and Eyring are the two workhorses, and the per-band spread between
them is itself a diagnostic. The diagram runs the room of this section
through both, with the validity boundary every statistical formula shares.

![Block diagram of the reverberation-time prediction: a 10 by 7 by 3.5 metre room with 245 cubic metres, 259 square metres and a mean absorption rising from 0.21 at 125 hertz to 0.51 at 4 kilohertz feeds the Sabine and Eyring formulas, whose per-octave-band table runs from 0.74 to 0.30 seconds for Sabine and 0.66 to 0.22 seconds for Eyring, Eyring reading 11 to 29 percent shorter; a closing note bounds the domain of validity to a diffuse field, excluding bands below the Schroeder frequency, coupled volumes and corridor-like rooms](https://docs.jmrp.io/phonometry/media/diagram_reverberation_prediction.svg)

*The same 245 m³ room through Sabine and Eyring alone. Eyring reads 11 %
shorter at 125 Hz and 29 % shorter at 4 kHz, because its correction grows with
the mean absorption, which here rises from 0.21 to 0.51 across the spectrum.
The closing note is the domain both share.*

## 4. Choosing a model, and when every model fails

The single axis along which the first three models disagree is the mean
absorption, so it is worth seeing them plotted against it before reading the
bullets below.

![Upper panel, reverberation time on a logarithmic axis against mean absorption from 0.02 to 0.99 for a 120 cubic metre shoebox: the Sabine, Eyring and Millington-Sette curves coincide at low absorption and separate above about 0.2, with Sabine flattening to a finite value at total absorption while Eyring and Millington-Sette fall to zero. Lower panel, the departure of Eyring and Millington-Sette from Sabine in per cent, passing minus 10 per cent at a mean absorption of 0.2 and minus 30 per cent at 0.5](https://docs.jmrp.io/phonometry/media/reverberation_model_absorption.svg)

*A 8 × 5 × 3 m shoebox with a uniform mean absorption swept from 0.02 to 0.99.
Below $\bar\alpha \approx 0.1$ all three coincide, which is why Sabine survived
a century of use. Eyring and Millington-Sette are identical for a uniform
distribution and fall away from Sabine as the absorption rises: −10 % at
$\bar\alpha$ = 0.2, −28 % at 0.5, −61 % at 0.9. At the right-hand edge the
structural difference is visible: Sabine still predicts a finite 0.12 s for a
room with an opening in every direction, while the logarithmic models reach
zero.*

**Show the code for this figure**

```python
# `room` and `plt` are the imports of the figure block at the top of the page.
import numpy as np

alpha = np.linspace(0.02, 0.99, 200)
volume, area = 120.0, 158.0
sab = np.array([room.sabine_reverberation_time(volume, [(area, a)]) for a in alpha])
eyr = np.array([room.eyring_reverberation_time(volume, [(area, a)]) for a in alpha])

fig, (top, bottom) = plt.subplots(2, 1, sharex=True)
top.semilogy(alpha, sab, label="Sabine")
top.semilogy(alpha, eyr, "--", label="Eyring")
top.set_ylabel("T [s]")
top.legend()
bottom.plot(alpha, 100.0 * (eyr / sab - 1.0))
bottom.set_xlabel(r"Mean absorption $\bar\alpha$")
bottom.set_ylabel("Departure from Sabine [%]")
plt.show()
```

The five formulae are not rivals on a single axis of accuracy; each has a
domain of validity:

- **Sabine** is the tool for live rooms with low, reasonably even
  absorption (mean $\bar\alpha$ up to roughly 0.2): classrooms, halls,
  reverberation chambers. It is also the convention wired into measurement
  practice, because the ISO 354 absorption coefficient is *defined* through
  Sabine's formula, so feeding reverberation-room data back into Sabine is
  self-consistent even where the formula is strained. Its structural defect
  shows at high absorption: with $\alpha = 1$ on every surface (an opening
  in every direction) it still predicts a finite reverberation time.
- **Eyring** is the choice for evenly treated rooms with substantial
  absorption: studios, treated offices, listening rooms. It reaches
  $T = 0$ for total absorption, and its correction over Sabine grows with
  $\bar\alpha$ (about 10 % shorter at $\bar\alpha = 0.2$, 30 % at 0.5).
- **Millington-Sette** handles a mix of very absorptive and hard surfaces
  better than a single mean, but it is meant for measured, sub-unity
  coefficients: a single surface with $\alpha_i = 1$ drives the whole
  prediction to zero. Reverberation-room coefficients at or above 1.0 are a
  documented ISO 354 outcome: edge diffraction scatters more energy into the
  sample than its flat area intercepts (see
  [Sound Absorption Measurement and Rating](https://docs.jmrp.io/phonometry/materials/absorbers/absorption-measurement/))
  and they lie outside the domain of the logarithmic term. To use
  Millington-Sette anyway you must bring the coefficient into $[0, 1)$, and
  that is a modelling decision the formula does not prescribe: whatever
  adjustment you choose (limiting to just below 1 is common), record it
  alongside the prediction.
- **Fitzroy** and **Arau-Puchades** target shoebox rooms whose absorption
  is concentrated on one axis, the typical office or dwelling with a soft
  floor and ceiling between hard walls. Arau's geometric mean tempers
  Fitzroy's known over-prediction when one wall pair is very reflective.

**What each model accepts.** phonometry enforces each formula's own domain, so
an out-of-domain coefficient raises a `ValueError` rather than returning a
meaningless number:

| Model | Accepts | Why |
|:---|:---|:---|
| Sabine | any $\alpha_i \geq 0$, including above 1 | the linear sum $\sum_i S_i\alpha_i$ stays finite |
| Eyring | individual $\alpha_i$ above 1, provided $\bar\alpha < 1$ | only the mean enters $\ln(1-\bar\alpha)$ |
| Millington-Sette | every $\alpha_i < 1$ | its per-surface $\ln(1-\alpha_i)$ diverges at 1 |
| Fitzroy | each wall-pair mean $< 1$ | the means are the inputs, and each enters a logarithm |
| Arau-Puchades | each wall-pair mean $< 1$ | same |

All five reject a coefficient above 2.0 outright, on the grounds that it is a
percentage passed as a fraction: measured ISO 354 values do not exceed about
1.2.

**When every formula fails.** All five inherit the same assumption: a
diffuse field, with sound arriving equally from all directions at every
point, that stays diffuse while it decays. The common breakages:

- [**Below the Schroeder frequency**](https://docs.jmrp.io/phonometry/buildings/rooms/room-image-sources/#3-modes-of-a-rectangular-room)
  the band holds a handful of discrete
  modes (the animation in §1) and a statistical reverberation time is not
  defined at all; each mode decays at its own rate set by the wall
  impedances it actually touches.
- **Coupled volumes** (a hall with an open stage house, two rooms through a
  doorway) produce double-slope decays; no single $T$ exists, and the
  measured T20 and T30 disagree (the
  [curvature diagnostic](https://docs.jmrp.io/phonometry/buildings/rooms/room-acoustics/#reading-edt-t20-and-t30-against-each-other)
  of the room-parameter guide).
- **Disproportionate rooms** (corridors, low flat halls) with the
  absorption on one surface pair keep a grazing sound field parallel to the
  hard surfaces that the absorber barely touches; the measured time can be
  up to twice any statistical prediction, the practical experience recorded
  in EN 12354-6 (see
  [Sound absorption in enclosed spaces](https://docs.jmrp.io/phonometry/buildings/rooms/enclosed-space-absorption/)).
  That grazing field is the striped survivor of the §2 clip, which measures
  it outliving the whole Sabine-Eyring band of its own section.
- **Focusing geometries** (domes, curved rear walls) concentrate late
  energy instead of mixing it, producing position-dependent decays no
  single-number formula can represent.

Scattering objects restore the mixing the models assume: a furnished room
follows the statistical prediction distinctly better than the same room
bare, beyond what the furniture's own absorption area accounts for. The clip
below is that mechanism with the absorption taken out of it, so only the
mixing is left: an 800 Hz wavefront enters a 4 m rigid-walled hall filled
with rigid columns 10 to 17 cm across, a quarter to two fifths of the
42.9 cm wavelength. Every column diffracts the front and sheds a scattered
wavelet, the wavelets interfere, and within a few passes the specular front
has become energy spread over the whole hall with no preferred direction,
which is the assumption Sabine and Eyring both start from, arriving here as
a result rather than as a hypothesis. Nothing in the hall absorbs, so what
you are watching is *only* the redistribution; the decay at the end is the
energy draining out through the two open ends.

**Video:** [A wavefront mixing in a hall of columns (2D FDTD)](https://docs.jmrp.io/phonometry/media/anim_fdtd_pillar_hall.webm): An 800 Hz plane wavefront sweeps a 4 m rigid-walled hall filled with a staggered colonnade of rigid columns 10 to 17 cm across, simulated at 2.5 mm; every column diffracts the front and sheds a scattered wavelet, and the wavelets interfere until the whole hall is filled with structured energy that then drains through the absorbing ends.

In practice, quote a *band* of predictions (Sabine and Eyring, or Fitzroy and
Arau-Puchades for axial cases) rather than a single value; where the models
spread, the room is telling you its field is not diffuse.

### What value to aim at

A prediction is only useful against a target, and a reverberation-time
requirement is a **range across the octave bands**, not a single upper limit,
which is why the fiche of §5 prints the target as a reference line and emits no
verdict. Anchors worth carrying:

- **Speech-critical rooms** (classrooms, meeting rooms, lecture theatres) land
  between roughly 0.4 s and 0.8 s at mid frequencies, tightening as the room
  gets smaller. ANSI/ASA S12.60-1:2010 Table 1 is the strictest widely used
  version of that: for unoccupied, furnished core learning spaces it caps the
  average of the 500 Hz, 1 kHz and 2 kHz octaves at **0.6 s** up to 283 m³ and
  **0.7 s** from 283 m³ to 566 m³, with no requirement above that, and it
  additionally asks the smaller rooms to be adaptable down to 0.3 s.
- **Open-plan offices, corridors and stairwells** are specified by absorption
  *area* rather than by reverberation time, because the field in them is too
  far from diffuse for a single $T$ to describe (this is EN 12354-6 clause
  4.5's own advice, and the ISO 3382-3 quantities of
  [Open-Plan Office Acoustics](https://docs.jmrp.io/phonometry/buildings/rooms/open-plan-acoustics/)
  replace $T$ entirely).
- **Music rehearsal and performance spaces** need more than 1 s and rise with
  volume; ISO 3382-1 Table A.1 quotes 1.0 s to 3.0 s as the typical EDT range
  of unoccupied concert and multi-purpose halls up to 25 000 m³.

Two shape rules go with the numbers. The target is normally read at the 500 Hz
and 1 kHz octaves, and the goal across frequency is a **flat** spectrum: a
low-frequency rise of more than about 20 % over the mid-frequency value is what
makes a treated room sound boomy even when it passes at 500 Hz, and a rise of
that size is tolerated in music spaces but not in classrooms. And because the
prediction band is often 20 to 30 % wide (the §0 classroom spans 0.59 s to
1.33 s), a design that only just meets its target on the most favourable model
has no margin at all.

## 5. Prediction report (`.report()`)

`ReverberationModelResult.report(path)` renders a one-page PDF fiche of the
prediction: a basis line marking it a **design-stage prediction** by the five
statistical-acoustics models, an optional metadata header block (client, room,
description, room volume, total surface area, climate), a per-band table with
one reverberation-time column per model beside the model comparison plot
(`.plot()`), and the boxed mid-frequency reverberation time from Arau-Puchades
(the recommended model for a non-uniform absorption distribution) with the
per-model spread alongside. It is a prediction, not a measurement: the five
models bracket the reverberation time likely to occur, so no PASS/FAIL verdict
is emitted. A target reverberation time supplied through the metadata's
`requirement` field is printed as a reference line only, since a room
reverberation time is a target range rather than a strictly
higher/lower-is-better quantity. It uses the same `ReportMetadata` container and
rendering engine as the other fiches; passing `metadata=None` produces a bare
prediction 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).

```python
from phonometry import ReportMetadata, room

result = room.reverberation_time_models(
    (8.0, 5.0, 3.0),                       # a shoebox room, one treated wall pair
    ([0.10, 0.15, 0.30, 0.45, 0.55, 0.60], # treated wall pair, per octave band
     [0.08, 0.10, 0.12, 0.15, 0.18, 0.20], # side walls
     [0.05, 0.08, 0.10, 0.12, 0.15, 0.18]),# floor/ceiling
    frequencies=[125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0],
)
result.report(
    "reverberation_fiche.pdf",
    metadata=ReportMetadata(
        specimen="Classroom, one wall lined with a broadband absorber",
        test_room="Classroom C1",
        temperature_c=20.0, relative_humidity_percent=50.0,
        laboratory="Phonometry Reference Laboratory",
        requirement=0.8,          # printed as a target reference line, no verdict
    ),
)                                 # the five-model table + the boxed T_mid
```

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

**Report:** [Reverberation-time prediction example report (PDF)](https://docs.jmrp.io/phonometry/media/reverberation_prediction_example.pdf): One-page reverberation-time prediction fiche: a metadata header (client, room, description, room volume, total surface area, temperature, humidity and pressure), the octave-band table with one reverberation-time column per model (Sabine, Eyring, Millington-Sette, Fitzroy and Arau-Puchades from 125 Hz to 4 kHz) beside the five-model comparison plot, the boxed mid-frequency reverberation time from Arau-Puchades with the per-model spread alongside, and a target reverberation-time reference line (no PASS/FAIL verdict).

Reverberation-time prediction fiche (ReverberationModelResult.report), the five-model table and the boxed T_mid.

## What this guide covers

- **Covered.** The classical statistical-acoustics reverberation-time formulae of Sabine
  (1922), Eyring (1930), Millington-Sette (1932), Fitzroy (1959) and
  Arau-Puchades (1988), each with its own domain-of-validity check and the ISO
  9613-1-derived air-absorption term, via `room.sabine_reverberation_time`,
  `room.eyring_reverberation_time`, `room.millington_sette_reverberation_time`,
  `room.fitzroy_reverberation_time`, `room.arau_puchades_reverberation_time` and
  the combined `room.reverberation_time_models`.

- **Not covered.** EN 12354-6's own Clause 4 model (a Sabine calculation with an added object
  term and the ISO 9613-1 air term) is a distinct, standard-defined calculation
  that lives in the [Sound absorption in enclosed spaces
  guide](https://docs.jmrp.io/phonometry/buildings/rooms/enclosed-space-absorption/), not here. None
  of the five formulae here models coupled volumes, disproportionate rooms or
  focusing geometries: their double-slope or position-dependent decays fall
  outside every diffuse-field assumption (§4). For a measured decay that shows
  those symptoms, use the T20/T30 curvature diagnostic of the [Room Acoustics
  guide](https://docs.jmrp.io/phonometry/buildings/rooms/room-acoustics/) rather than a statistical
  prediction.

## See also

- [Sound Absorption Measurement and Rating](https://docs.jmrp.io/phonometry/materials/absorbers/absorption-measurement/):
  the ISO 354 measurement every $\alpha_\mathrm{s}$ on this page comes from, its mounting
  dependence, and the ISO 11654 rating that must *not* be substituted for the
  per-band table (§0).
- [Room acoustic parameters (ISO 3382-1/2)](https://docs.jmrp.io/phonometry/buildings/rooms/room-acoustics/#how-much-decay-range-the-noise-floor-allows):
  the measured counterpart, and the curvature diagnostic that tells you a
  prediction's diffuse-field assumption has failed.
- [Modes of a rectangular room](https://docs.jmrp.io/phonometry/buildings/rooms/room-image-sources/#3-modes-of-a-rectangular-room):
  what is below the Schroeder frequency, where none of these five formulae
  applies.
- API reference: [`room.reverberation_prediction`](https://docs.jmrp.io/phonometry/reference/api/rooms/reverberation-prediction/) and [`environment.propagation.air_absorption`](https://docs.jmrp.io/phonometry/reference/api/environment/air-absorption/).
- Theory: [Sound insulation and absorption, predicted](https://docs.jmrp.io/phonometry/reference/theory/rooms-buildings/#sound-insulation-and-absorption-predicted-en-12354-1-2-6-bies-cremer-hopkins): the diffuse-field assumption every Sabine-family formula makes, and where it stops holding.
- [Published catalogues](https://docs.jmrp.io/phonometry/reference/catalogues/absorption-coefficients/): the
  band-by-band absorption coefficients the books print, with the mounting each
  one was measured on.

## References

- **Arau-Puchades, H. (1988).** [An improved reverberation formula](https://www.ingentaconnect.com/content/dav/aaua/1988/00000065/00000004/art00003). *Acustica*, *65*(4), 163-180.
  The geometric-mean combination of §2 (its Formula 18). The linked record is the publisher page at Ingenta.
- **Carrión Isbert, A. (1998).** [*Diseño acústico de espacios arquitectónicos*](https://openlibrary.org/books/OL23159935M). Edicions UPC.
  A Spanish-language textbook treatment of the reverberation models and their use in room design. ISBN 978-84-8301-252-9.
- **European Committee for Standardization. (2003).** [*Building acoustics — Estimation of acoustic performance of buildings from the performance of elements — Part 6: Sound absorption in enclosed spaces*](https://knowledge.bsigroup.com/products/building-acoustics-estimation-of-acoustic-performance-of-buildings-from-the-performance-of-elements-sound-absorption-in-enclosed-spaces) (EN 12354-6:2003).
  The classical formulae predate the normative world and enter it through this standard, whose Clause 4 model is a Sabine calculation with object and air terms; see the enclosed-space absorption guide. The linked catalogue record is the BSI Knowledge page for BS EN 12354-6:2003.
- **Everest, F. A. (2001).** [*Master handbook of acoustics*](https://openlibrary.org/isbn/9780071360975) (4th ed.). McGraw-Hill.
  The Fig. 7-22 worked example the conformance suite reproduces: Example 1, an untreated 23.3 × 16 × 10 ft room, whose six printed Sabine reverberation times the SI implementation reproduces to ≤ 0.02 s, reinforced by hand-computed closed-form values and the model identities (every model collapses to Eyring for uniform absorption; Eyring collapses to Sabine as the absorption tends to zero), which transitively carry that real-data anchor to the whole family. ISBN 978-0-07-136097-5.
- **Eyring, C. F. (1930).** Reverberation time in "dead" rooms. *The Journal of the Acoustical Society of America*, *1*(2A), 217-241. https://doi.org/10.1121/1.1915175
  The mean-free-path derivation behind the −S ln(1−ᾱ) term of §1.
- **Fitzroy, D. (1959).** Reverberation formula which seems to be more accurate with nonuniform distribution of absorption. *The Journal of the Acoustical Society of America*, *31*(7), 893-897. https://doi.org/10.1121/1.1907814
  The axial split into three wall-pair decays of §2.
- **International Organization for Standardization. (1993).** [*Acoustics — Attenuation of sound during propagation outdoors — Part 1: Calculation of the absorption of sound by the atmosphere*](https://www.iso.org/standard/17426.html) (ISO 9613-1:1993).
  The atmospheric-absorption coefficient behind the air term 4mV; see the outdoor propagation guide.
- **International Organization for Standardization. (2003).** [*Acoustics — Measurement of sound absorption in a reverberation room*](https://www.iso.org/standard/34545.html) (ISO 354:2003).
  Defines the measured absorption coefficient via Sabine's formula, the self-consistency argument of §4.
- **Kuttruff, H. (2016).** *Room acoustics* (6th ed.). CRC Press. https://doi.org/10.1201/9781315372150
  The diffuse-field theory, its limits and the modern assessment of the classical formulae behind §4.
- **Millington, G. (1932).** A modified formula for reverberation. *The Journal of the Acoustical Society of America*, *4*(1), 69-82. https://doi.org/10.1121/1.1915588
  The per-surface logarithmic absorption term of §1.
- **Sabine, W. C. (1922).** [*Collected papers on acoustics*](https://archive.org/details/collectedpaperso00sabi). Harvard University Press.
  The original reverberation experiments and the T = 0.161 V/A law of §1. The linked copy is the free scan at the Internet Archive.
