Classification of Reverberant Acoustic Situations

Jens Schroder1 , Thomas Rohdenburg2 , Volker Hohmann1 , Stephan D. Ewert1

1

University of Oldenburg, Institute of Physics, Medical Physics, Germany,

Email: [email protected]; [email protected]
Fraunhofer Institute for Digital Media Technology, 26129 Oldenburg, Germany

Introduction
In daily communication, speech intelligibility depends on
the acoustic surrounding or acoustic situation. Particularly for hearing impaired persons, speech understanding
is often problematic if speech is distorted by (room)
reverb, noise or competing talkers. Acoustic situations
are characterized by different dominating types of distortion. Hearing aids might provide appropriate algorithms
to enhance speech intelligibility in the different acoustic
situations. A robust and fast automatic classification
of the acoustic situation should therefore select the
appropriate hearing aid algorithm without requiring an
action of the hearing aid wearer. This study is concerned
with the automatic estimation of the reverberation time
(T 60) in natural situations and with unknown excitation
signal.
Acoustic test situations were generated by
convolving speech signals with artificial and real room
impulse responses with T 60 times ranging from 0.05
to 4 s. Features derived from the cepstral mean, the
autocorrelation function and from the distribution of
modulation energy were used to blindly estimate different
reverb times.

Impulse Response Model

In natural environments sound is often received as a
superposition of direct and reflected sound from walls or
objects in a room. Direct, early reflexions arrive first at
the ear and after multiple reflexions and superpositions of
reflexions from many objects they are diffuse and called
reverberation. The level of reflexions in a room impulse
response decreases due to attenuation effects and scattering. This decay is often assumed to be nearly exponential
([1],[6]). Thus, a reverberant impulse response can be
approximated by an exponentially decaying part and if
measurement noise or background noise occured by a
constant part [1]:
h(t) = Aexp et/ n1 (t) + Anoise n2 (t),

(1)

where Aexp und Anoise are scalar, is the decay parameter in seconds, t is the time in seconds and n1 (t) and
n2 (t) present two independent noise processes.
A common measure for reverberation is the time until
the impulse response has decreased by 60 dB. In (1), the
reverberation time, T 60, can be calculated directly from
the decay parameter :
T 60 = ln(103 ) 6.908 .

(2)

To calculate the T 60 time from a measured impulse

response different solutions exist. A procedure suggested

in [1] is to fit the power by a least square fit. From

equation (1) the instantaneous power can be derived as
q
(3)
a(t) = A2exp e2t/ + A2noise .
The parameters Aexp , and Anoise are evaluated by a
least square fit of the form
Z
2
min [as (t) y s (t)] dt
(4)
where s = 0.5 is a scaling factor to improve the results
[1]. The T 60 time was then calculated as stated in (2).

Blind Estimation Procedures

The goal of this study is to estimate the T 60 time from
a reverberated speech sample without having explicit
information about neither the impulse response nor the
underlying clean speech.
Three different methods are used in the following and
their results are compared.

Cepstral Mean
To estimate the impulse response from an unknown
reverberated signal there exists the theory of blind
homomorphic deconvolution [3], [4], [5]. Here a reverberated speech signal is assumed as:
sir(t) = s(t) h(t),

(5)

where denotes the convolution product, s(t) clean

speech and h(t) the impulse response. A Fourier transformation turns the convolution into a multiplication.
The logarithm transforms this product into a sum. A
further inverse Fourier transformation converts the sum
into the cepstral domain where the additivity is preserved
(see figure 1). If it is assumed that the cepstrum
F

log

F 1

b ) + H(f
b ) sb(q) + b
s(t) h(t) S(f ) H(f ) S(f
h(q)
Figure 1: Calculation of the cepstrum from a convoluted
input signal

of the clean speech is nearly uncorrelated in adjacent

windows, averaging over a few cepstra estimates the
mean cepstrum of the impulse response. By inversing
h(t) can be derived.
To keep the inverse cepstrum stable and causal, only the
minimum phase part of the deconvolved impulse response
is taken [3]. The whole deconvolution scheme by cepstral
mean is shown in figure 2.
The T 60 time can then
be estimated from the resulting impulse response by the
above mentioned least square fit.

cepstrogram

|F( )|

filtering
(causal)

F (log( ))
F -1(log( ))

1 S( )
T
T

F -1(exp(F( )))

t=1

time

|F( )|

inverse
cepstrum

Stimuli and Methods

level

|F( )|

mean

-1

F -1(log( ))

time

Figure 2: Blind deconvolution of reverberated speech by

cepstral mean estimation of the minimum phase part of the
impulse response

Autocorrelation
The autocorrelation function Rsir,sir (t) of a reverberant
signal sir(t) is the convolution product of the autocorrelation functions of the underlying clean speech s(t) and
the room impulse response h(t).
Rsir,sir (t)

= s(t) h(t) s(t) h(t)

= Rs,s (t) Rh,h (t).

(6)

If the clean speech is considered to have a peaky autocorrelation function then the following approximation is
possible:
Rsir,sir (t) Rh,h (t).
(7)
For an exponential function, the autocorrelation function
for positive times has the same exponential decay parameter . Thus we assume the autocorrelation function
of the reverberated signal to decay like the underlying
impulse response. To reduce estimation errors averaging
over overlapping windows was performed. From the
averaged autocorrelation function the T 60 time was
estimated due to equation (4).

Speech to Reverberation modulation energy ratio (SRMR)

Typically, clean speech shows the strongest modulation energy at a modulation frequency of about 4 Hz.
The distribution of modulation energy shifts to higher
modulation frequencies for reverberated speech as a
consequence of the whitening effect by the impulse
response which is considered to be a damped gaussian
white noise. The higher the reverberation time, the
more modulation energy occurs at high modulation
frequencies. Thus [7] suggested a comparison of energies
at high and low modulation frequencies. Here, the
modulation spectrogram was seperated into low and
high modulation frequency regions and the ratio of the
respective energies was calculated. This ratio is called
SRMR (Speech to Reverberation Modulation energy
Ratio).
In comparision to [7], the weighting of the high and
modulation
spectrogram

weighting

SRMR
G,K

high

low
time

low

SE
SE

g,k=1
G,K
g,k=1

To analyse the three estimation methods, clean speech

material of four speaker sets sampled at 16 kHz was
used: Two male, German speakers, one female, German
speaker and a set of female, English speakers.
The clean speech material was reverberated by convolution with room impulse responses.
To characterize the impulse responses, the T 60 times
were estimated from the impulse responses using the
method described in (4) and are referred to as real T 60
times in the following.
For the first test setup, referred to as artifical impulse
response setup (artificial IR setup), impulse responses
were generated by source imaging [2] to achieve a wide
range of T 60 times with controlable spectral (white)
properties. For these impulse responses a rectangular
room was simulated and its size and reflexion coefficients
were adjusted to produce T 60 times ranging from 50 ms
to 4 s, increasing at a factor of two. The sound source
and the receiver were distributed randomly in the room
and the positions differed between the impulse responses.
Independent of the T60 time the length of each impulse
response was 2 s with 16 kHz sampling frequency.
All reverberant speech samples were derived from the
same underlying clean speech. In this paper, the results
for one of the male German speakers are shown.
The second test setup, called real impulse response
setup (real IR setup), consisted of real impulse responses
selected from a commercial impulse response library.
Three groups of T 60 times were defined: a dry one
(0.16 - 0.36 s) with a mean T 60 = 0.3 s, a medium
reverberated one (0.72 - 1.02 s) with a mean T 60 =
0.9 s and a reverberant one (1.71 - 1.98 s) with a mean
T 60 = 1.9 s. Each group consisted of four impulse
responses. Each of the four impulse responses of a group
are convolved with different speech material from one of
the four speaker sets.
To test all three features, an estimation of the T 60 time
respectively the SRMR was done every 0.5 s for a total
time of 100 seconds for the artificial IR setup and 40
seconds for each speaker/impulse response of a T 60 group
for the real IR setup.
For the cepstral mean and the autocorrelation feature
the window lengths were 0.05 s, 0.2 s, 0.5 s, 1 s, 2 s,
3 s, and 4 s. The overlap between two windows was
7/8 and the averaging time 5 times the window length
(=
b 33 windows). The first 6.3 ms (=
b 100 time samples at
16 kHz sampling frequency) of the deconvolved impulse
responses were skipped for the fitting.
The window length for the SRMR feature was 1 s.

Results

g,k

high
g,k

Figure 3: Algorithm for the estimation of SRMR

low modulation frequency regions was changed. Low

modulation frequencies were defined to be smaller than
28.9 Hz or smaller than 10% of the corresponding
bandwidth of the auditory filter.

Cepstral Mean
The means of 200 T 60-time estimates for the artificial
IR setup are plotted in Fig. 4 (solid lines) as a function
of the analysis window duration for three different real
T 60 times indicated by the dotted lines.
The T 60time estimates depend on the analysis window duration,

blind estimated T60s with different window lengths

means of valid T60s (double logarithmic plot)

means of valid T60s

4.5

3.5

3.2
1.6
estimated T60 /s

3
estimated T60 /s

real T60: 0.05 s

real T60: 0.1 s
real T60: 0.2 s
real T60: 0.4 s
real T60: 0.8 s
real T60: 1.6 s
real T60: 3.2 s

2.5
2
1.5
1

0.8

0.2

66

53.8

1.9

96.5
100

100
97

94.1
0.9

0.1

100

0.5

0.3

0.05
0

0.5

1.5

2
2.5
3
window length /s

3.5

4.5

Figure 4: Cepstral mean: Means and standard deviations

of 200 estimated T 60 times per window length and impulse
response of the artificial IR set up (solid lines). The dotted
lines of same color represent the real T 60 time.

starting at small values for short window durations and

asymptoting against the real T 60 times with increasing
window duration. The T 60-time curves flatten off at a
window duration corresponding to about two times real
T 60 time. Since the estimates depend on the window
duration, the proper window for a good T 60-time estimation has to be chosen blindly. To do so, the algorithm
suggested here successively calculates T 60-time estimates
for increasing analysis window durations. Following the
slope analysis given above, the validity of the estimates
is judged blindly by monitoring the differences of the
T 60-time estimates derived for two succesive analysis
window durations. If the slope calculated from the two
last estimates is smaller than an empirically adjusted
criterion of 0.1, the last estimate is considered to be valid.
A second criterion to judge the validity of the estimate is
to monitor the ratio between the energy of the fitted noise
and the energy in the fittet exponential decay calculated
over the duration of the current T 60-time estimate:
PT 60 2
N
Anoise
=
.
(8)
PT 60
2
S
(Aexp exp( t))

Again, an empirically adjusted criterion of N

S < 0.25 has
to be met in order to judge the estimate as valid. If
both criterions are met, a valid T 60-time estimate was
calculated by the algorithm.
The means of the valid T 60-time estimates are plotted
in Fig. 5 as a function of the real T 60 times. The left
panel shows the results for the artificial IR setup and
the right panel for the real IR setup.
The numbers
indicate the percentage of T 60-time estimates which
satisfied both validity criterions. Apparently, a lower
limit for the estimated T 60 times exists at about 200 ms
for small real T 60 times. For longer real T 60 times,
the estimates match very well. The number of valid
T 60 times decreases since longer real T 60 times need
longer window duration for a proper estimation. The
calculation time for a valid estimate is about ten times
the real T 60 time: the best window duration is about
two times the real T 60 time and the averaging time is
five times the window length. Although the computation
time for the very accurate T 60-time estimate appears
quite long, the successive prolonging of the window
duration in the algorithm allows an early estimation of

0.05

0.1

0.2

0.4
0.8
real T60 /s

1.6

3.2

0.3

0.9
real T60 /s

1.9

Figure 5: Cepstral mean: Means and standard deviations

of valid T 60-time estimates. The numbers indicate the
percentage of the valid estimates in relation to all estimates.
Left panel: artificial IR set up, 200 estimations per impulse
response. Right panel: real IR set up, 320 estimations per
impulse response group

the lowest possible value for the currently observed T 60

time: The longer the actual window duration is, the
longer the T 60 time will be.

Autocorrelation
For the autocorrelation feature, the same sound material
and paramters as for the cepstral mean feature were used
(see above).
The means of 200 T 60-time estimates per window and
impulse response of the articicial IR setup are plotted
in Figure 6, comparable to Figure 4. Comparable to
blind estimated T60s with different window lengths
4.5
real T60: 0.1 s
real T60: 1.6 s
real T60: 3.2 s

4
3.5
3
estimated T60 /s

74

0.4

0.5

real T60: 0.3 s

real T60: 0.9 s
real T60: 1.9 s

10.5
estimated T60 /s

real T60: 0.1 s

real T60: 1.6 s
real T60: 3.2 s

2.5
2
1.5
1
0.5
0
0.5
1

0.5

1.5

2
2.5
3
window length /s

3.5

4.5

Figure 6: Autocorrelation: The estimated T 60 times

as means with standard deviations of 200 estimations per
window and impulse response of the articicial IR setup (solid
lines). The real T 60 times are indicated as dotted lines of the
same color.

the cepstral mean feature, the estimated T 60 times of

the autocorrelation feature asymptote against the real
T 60, though the deviations are much higher here than
in case of the cepstral mean feature. The prolonging
of the analysis window duration was again used and
valid estimates were derived when the same slope and
N
S criteria as in case of the cepstral mean feature were
met. The results for the valid T 60-time estimates are
shown in Figure 7.
The results are similar to those
from the cepstral mean. Again, there is a lower limit at
about 200 ms for the T 60-time estimates. Above 200 ms
the means of estimated T 60 times are matching the real
ones well, though the deviations are larger than those of
cepstral mean. The number of valid T 60 times decreases
heavily for all impulse responses.

means of valid T60s (double logarithmic plot)

3.2

estimated T60 /s

1.6
0.8

means of valid T60s

0.5
15

20.5 48.5

20.6

1.9

19.5
33

0.4
0.2

real T60: 0.3 s

real T60: 0.9 s
real T60: 1.9 s

estimated T60 /s

real T60: 0.05 s

real T60: 0.1 s
real T60: 0.2 s
real T60: 0.4 s
real T60: 0.8 s
real T60: 1.6 s
real T60: 3.2 s

70.5

35.9

0.9

0.1

91.6
0.3

0.05
0.05

0.1

0.2

0.4
0.8
real T60 /s

1.6

3.2

0.3

0.9
real T60 /s

1.9

Figure 7: Autocorrelation: Means and standard deviations

of valid T 60-time estimates for the different impulse
responses. The numbers show the percentage of valid T 60time estimates in relation to all estimates. Left panel:
artificial IR setup, 200 estimations per impulse response; right
panel: real IR setup, 320 estimations per impulse response
group.

Speech to Reverberation modulation energy ratio (SRMR)

In Figure 8 the means and standard deviations of SRMRs
are plotted.
The left panel shows the results for the artificial IR setup.
means (semi logarithmic plot)

0.8

0.6

1.6

0.4

real T60: 0.3 s

real T60: 0.9 s
real T60: 1.9 s

1.4
1.2
SRMR

SRMR

means

real T60: 0.05 s

real T60: 0.1 s
real T60: 0.2 s
real T60: 0.4 s
real T60: 0.8 s
real T60: 1.6 s
real T60: 3.2 s

0.6
0.4
0.2

0.05

0.1

0.2

0.4
0.8
real T60 /s

1.6

3.2

Acknowledgements
This work was supported by the Bundesministerium f
ur
Bildung und Forschung (BMBF) project Modellbasierte
Horsysteme.

References
[1] M. Karjalainen, P. Antsalo, A. Makivirta, T.
Peltonen, V. Valimaki, Estimation of Modal Decay
Parameters from Noisy Response Measurments, J.
Audio. Eng. Soc., Vol. 50, No. 11, pp. 867-878, Nov.
2002

0.8

0.2

shorter correlation duration like white Gaussian noise.

Another observation is that the number of valid T 60-time
estimates drops towards longer real T 60 times with the
current choice of the longest analysis window duration
of 4 s, particularly for the autocorrelation feature. The
use of even longer analysis window duration seems not
feasible with the practical application in mind. The
calculation times that where achieved are about 10 times
the T 60 time (window length = 2 T 60, averaging time
= 5 window length). Nevertheless an early classification
into short and long T60 times is possible by monitoring
the active window duration in the algortihm: The larger
the active window (even if the T 60 time is not valid) the
longer the T 60 time.
For the SRMR feature, the reliable results from [7] could
not be reproduced. By averaging over some SRMRs or
modulation spectrograms this feature might, however,
be usefull in combination with the cepstral mean or
autocorrelation feature.
In the next step all three features are combined in a
classifier with a gaussian mixture model (GMM) for
robustness.

0.3

0.9
1.9
real T60 /s

Figure 8: Means and standard deviations of calculated

SRMRs per impulse response over the real T 60 times. Left
panel: artificial IR setup, 200 estimations per impulse
response; right panel: real IR setup, 320 estimations per
impulse response group.

It is obvious that the SRMRs between 50 and 400 ms

are nearly equal and that SRMRs decrease beyond about
400 ms real T 60 time. The large standard deviations
indicate that even a meaningful classification in rough
T 60-time categories would fail without averaging over a
large number of SRMRs.

Conclusions
Three different methods for the estimation of the reverberation time T 60 were presented. It was shown that
for the cepstral mean and the autocorrelation feature
an estimation of the T 60 time via the N
S and slope
criteria is possible with very good accuracy above about
200 ms. The lower limit of estimated T 60 times at
about 200 ms is most likely related to the statistical
features of speech. Both methods assume that the speech
signal is statistically independent in successive time
windows which is not the case. Shorter T 60 times could
be only estimated with a input signal of significantly

[2] J. B. Allen, D. A. Berkley , Image method

for efficiently simulating small-room acoustics, J.
Acoust. Soc. Am., Vol. 65, No. 4, pp. 943-950, Apr.
1979
[3] A. Baskind, O. Warusfel, Monaural and binaural
processing for automatic estimation of room acoustics
perceptual attributes, IRCAM meeting, Sep. 2001
[4] A. V. Oppenheim, R. W. Schafer, Zeitdiskrete
Signalverarbeitung, Oldenbourg Verlag, 3. Auflage,
1999
[5] T. G. Stockham, Jr., T. M. Cannon, R. B.
Ingebretsen, Blind Deconvolution Through Digital
Signal Processing, Proceedings of the IEEE, Vol. 63,
No. 4, April 1975
[6] M. Hansen, A Method for Calculating Reverberation
Time from Musical Signals, Technical Report 60,
The Acoustic Laboratory, Technical University of
Denmark
[7] T. H. Falk, W.-Y. Chan, A Non-Intrusive Quality
Measure of Dereverberated Speech, Intl. Workshop
for Acoustic Echo and Noise Control, 2008