EMPLOYING LAPLACIAN-GAUSSIAN DENSITIES FOR SPEECH ENHANCEMENT

Saeed Gazor
Department of Electrical and Computer Engineering, Queens University, Kingston, Ontario K7L 3N6, Canada
AbstractA new efcient Speech Enhancement Algorithm (SEA)
is developed in this paper. A noisy speech is rst decorrelated and
then the clean speech components are estimated from the decorre-
lated noisy speech samples. The distributions of clean speech and
noise are assumed to be Laplacian and Gaussian, respectively. The
clean speech components are estimated either by MaximumLikeli-
hood (ML) or Minimum-Mean-Square-Error (MMSE) estimators.
These estimators require some statistical parameters that are adap-
tively extracted by the ML approach during the active speech or
silence intervals, respectively. A Voice Activity Detector (VAD) is
employed to detect whether the speech is active or not. The simu-
lation results show that this SEA performs as well as a recent high
efciency SEA that employs the Wiener lter. The complexity of
this algorithm is very low compared with existing SEAs.
1. INTRODUCTION
Most of Speech Enhancement (SE) research has focused on re-
moving the corrupting noise. Usually it is assumed that speech is
degraded by additive noise which is independent of clean speech.
In early implementations, spectral subtraction approach was widely
used. This approach estimates the Power Spectral Density (PSD)
of a clean signal by subtracting the short-time PSD of the noise
from the PSD of the noisy signal [1]. The Wiener and Kalman
lters have been used for SE [3, 4]. The noisy speech is used to
estimate an optimum lter adaptively, under the assumption that
speech and noise are Gaussian, independent and have zero mean.
Recently a signal subspace speech enhancement framework has
been developed (see [57] and references therein). This signal
subspace SE system decomposes the noisy signal into uncorrelated
components by applying the Karhunen-Lo` eve Transform (KLT).
The recent statistical modelling presented in [8] concludes that
the clean speech components, in decorrelated domains (e.g., in the
KLT and the DCT domains) as random variables have Laplacian
distributions, and noise components are accurately modelled by
Gaussian distributions. Therefore, the speech decorrelated com-
ponents could be accurately modelled as a multivariate Laplacian
random vector, while for noise a multivariate Gaussian model is
accurate. Based on these assumptions, we design a Bayesian SE
system to estimate the clean speech signal components.
This paper is organized as follows. Section 2 reviews the basic
principle of the decorrelation of speech signals. Section 3 provides
the statistical modelling that will be used in this paper. A SEA is
proposed in Section 4. The performance evaluation and conclusion
are summarized in Section 5. Section 6 is the conclusion.
2. DECORRELATION OF SPEECH SIGNALS
Let x(t) be the clean speech and the vector of samples of x(t) be
denoted by X(m) = [x(m), x(m1), , x(mK + 1)]
T
where (.)
T
is the transpose operation. Also, let Y (m) = X(m) +
N(m) denote the corresponding K-dimensional vector of noisy
speech, assuming that the noise vector N(m) is additive. Ap-
plying a linear transformation to the noisy the speech signal we
may approximately assume that signal components are uncorre-
lated in the transformed domain. Since the correlation between
speech signals is commonly rather high, a speech data vector can
be represented with a small error by a small number of compo-
nents [7]. In this paper, the speech signals are transformed into
uncorrelated components by using the DCT or the AKLT [6]. It
can be easily seen that v
i(m) = si(m) + ui(m), where vi(m),
si(m) and ui(m) are transformed components of Y (m), X(m)
and N(m), respectively. In order to develop our SEA, we further
assume that the uncorrelated components, i.e., {si, ui}
K
i=1
, are in-
dependent [8]. This assumption is not required if random variables
uncorrelated are Gaussian variables and are uncorrelated. As the
KLT is complex to compute, harmonic transforms such as Discrete
Cosine Transform (DCT) and Discrete Fourier Transform (DFT)
are used as suboptimal alternatives. Another motivation for using
these transforms instead of the AKLT is to avoid the subspace vari-
ations and errors of the AKLT [6]. Among the DCT and DFT, the
DCT is cheaper and reduces better the correlation of the signal and
compacts the energy of a signal block into some of the coefcients.
3. STATISTICAL MODELLING
The noise samples can be separated during silence intervals us-
ing a VAD. For a Gaussian noise component u
i(m) with variance

2
i
(m), we have: fu
i,m
(ui(m)) =
exp(
u
2
i
(m)
2
2
i
(m)
)

2
2
i
(m)
. If the sam-
ples {ui(i)}
m
i=mM
N
+1
are iid, the ML estimate of
2
i
is

2
i
=
1
M
N

m
t=mM
N
+1
|ui(t)|
2
. A lower-complexity estimator is

2
i
(m) = N

2
i
(m1) + (1 N)|ui(m)|
2
, (1)
where N is chosen to let the time constant of the above lter be
0.5 second, assuming that the variation of the noise spectrum is
negligible over a time interval of 0.5 second.
Assuming that the components of the clean speech in the decor-
related domains, {si(m)}
K
i=1
, follow zero-mean Laplacian are un-
correlated, we have fs
i,m
(si(m)) =
1
2a
i
(m)
e

|s
i
(m)|
a
i
(m)
, where ai(m)
is the Laplacian factor [8]. Similarly, the ML estimate of the
Laplacian Factor ai yields ai(m) =
1
M
S
m

t=mM
S
+1
|si(t)|. Sim-
ilarly, we use the following low-complexity substitute:
ai(m) = S ai(m1) + (1 S) |si(m)| . (2)
In our simulations, S is chosen to let the time constant of the
above adaptive process be 10msec, because the speech signal can
I - 297 0-7803-8484-9/04/$20.00 2004 IEEE ICASSP 2004

-
6
s

2
a

2
a ML
MMSE
v

Fig. 1. Comparison of MMSE and ML estimators of the clean

speech component s from the noisy input component v = s + u.
be assumed to be stationary over such a period. The estimation of
a
i in (2) is equivalent to the expected value of |si(m)|. Note that
there is no access to the clean signal, si(m). If the noise power is
small, we may use |vi(m)| as an approximation for |si(m)| in (2).
4. ESTIMATION OF CLEAN SPEECH COMPONENTS
In this section for simplicity of notation, we drop the time index
m and eigenvector index i. Here, the problem is to estimate the
clean speech component s when the noisy speech component v =
s+u is given. Assuming that the speech is detected as present and
the speech s and noise u components are independent, the joint
distribution of s and v is given by
f
s,v
(s, v) =
1
2a

2
2
e

|s|
a

|vs|
2
2
2
, (3)
and the conditional distribution of s given v is
f
s|v
(s|v) =
fs,v(s, v)
_
+
s=
fs,v(s, v) ds
. (4)
The Minimum Mean Square Error (MMSE) Estimator is the
conditional mean of s with f
s|v
(s|v) as the pdf. Using (3), the
MMSE estimator of the clean speech component s, is given as a
non-linear function of three inputs: 1) noisy speech component v,
2) noise variance
2
and 3) speech Laplacian factor a:
MMSE : s E {s|v} =
_
+

sf
s|v
(s|v) ds (5)
= ae

( + )e

erfc
_
+

2
_
( )e

erfc
_

2
_
e

erfc
_
+

2
_
+ e

erfc
_

2
_

where =
v
a
, =

2
i
a
2
i
and erfc(x) =
2

x
e
t
2
dt.
The ML Estimator of s given the observation v is the value for
which the likelihood function f
v|s
(v|s) is maximum. Maximizing
f
v|s
(v|s) is equivalent to maximizing (3). Thus, we have
ML : s arg max
s
f
v|s
(v|s) = arg max
s
fs,v(s, v) (6)
= arg min
s
_
|s|
a
+
|v s|
2
2
2
_
=

v

2
a
, if v

2
a
,
0, if |v|

2
a
,
v +

2
a
, if v

2
a
.
Figure 1 depicts the MMSE and ML estimators (5) and (6) ver-
sus the noisy input signal v for a given value of

2
a
. We nd
that these estimators operate very similarly; if the amplitude of
the noisy input is large (i.e, |v| 2

2
a
), then the magnitude of
Noisy
Signal
Voice Activity
Detector
-
Analog-to-digital
and
Serial-to-parallel
y(t)
VAD Output
ai
(m),
2
i
(m)
Y (m)
Signal
Decorrelation
e.g., DCT
V (m)
Parameters
Estimation
Section 3
a
i(m)
?

2
i
(m)
?

-
-
?
?
Clean Speech Estimation

?
?

S(m)
P/S and D/A
x(t)
-

X(m)
-
Enhanced
Signal
Inverse
Transform
e.g., IDCT
Fig. 2. The block diagram of the proposed SE system.
the output is almost equal to the magnitude of the input minus

2
a
,
i.e, | s| max
_
0, |v|

2
a
_
. If the magnitude of the input v
is smaller than

2
a
the ML interprets it as noise and projects it to
zero, while MMSE attenuates the input.
Figure 2 and Table 2 is a summarize the proposed SEA. We
use the VAD in [9], since its design structure and assumptions are
as same as those of this paper; therefore, the transformation and
the estimation of the parameters could be shared between the pro-
posed SE and this VAD. Providing the speech and noise statistic
parameters, ai(m) and
2
i
(m), each DCT component will be used
to estimate the clean speech component along the corresponding
eigenvector using either ML estimation (5) or MMSE estimation
(6). The enhanced signal is obtained via the inverse transform of

S(m). The S/P is a buffer that takes the new portion of each new
input vector and feeds it to a shift register in order to produce the
stream of the enhanced signal x(t).
5. PERFORMANCE EVALUATION
The compromise between signal distortion and the level of resid-
ual noise is a well known problem in SE [3, 5]. In this section,
the performance of the proposed SEA is evaluated using objective
criteria, such as noise reduction criterion and distortion criterion.
The sampling frequency of the noisy speech signal is 11,025
Hz. The vector length, K, is chosen to be 80 samples, which
corresponds to approximately 7msec. The overlap between sig-
nal vectors is set at 70 samples. The overlap can be reduced to
reduce the computation complexity, at the expense of some per-
formance degradation. In this case, at each iteration 10 samples
of the enhanced signal is updated. This represents about 1msec
of the signal. Further reduction of this updating time interval pro-
vides only a very slight improvement. Software generated white
Gaussian noise, computer fan noise and lab noise are added to the
original speech signal with different SNRs, namely 0, 5, and 10dB.
The computer fan noise is picked up by a microphone, as is the lab
noise which is mainly the sound of a network switch.
Time Domain and Spectrogram Evaluation: First, the results
of the proposed SE algorithm are evaluated and compared with
the SEA in [6] in the time domain and the frequency domain by
means of the spectrogram. Figure 3 shows the results of enhanced
speech corrupted by a 5dB white noise. From this gure we ob-
serve that the enhanced speech has a lower noise level in the time
I - 298

Table 1. Summary of the soft VAD algorithm in [9]
Initialize:
S
= 0.913,
N
= 0.983, P
1|0
=
1
2
,
For each time step m do: v
i
(m) = [dct{Y (m)}]
i
,
For i= 1, 2, , K, if P
m|m1
0.5 do
a
i
(m) =
S
a
i
(m1) + (1
S
) |v
i
(m)|
if else do
2
i
(m) = N
2
i
(m1) + (1 N)u
2
i
(m) end;
f
0i(m) =
1
_
2
2
i
(m)
e

v
2
i
(m)
2
2
i
(m)
f1i(m) =
e

i,m
2
4a
i
(m)
_
e

i,m
erfc
_
i,m + i,m
_
2i,m
_
+
+e

i,m
erfc
_

i,m i,m
_
2
i,m
__
,
where
i,m
=
v
i(m)
ai(m)
, and
i,m
=

2
i
(m)
a
2
i
(m)
end;
L(m) =
K

i=1
f1i(m)
f0i(m)
P
m|m
=
L(m)P
m|m1
L(m)P
m|m1
+
_
1 P
m|m1
_ ,
P
m+1|m
= 01
_
1 P
m|m
_
+
11P
m|m
.
end;
Table 2. The structure of the proposed SEA.
Initialization: di(0) = 0, S = 0.913, N = 0.983,

S
and
N
are chosen to let the time constants of lters to
be 10msec and 0.5sec, respectively.
For each time step m do
V (m)=[v
1
(m), v
2
(m), , v
K
(m)]
T
dct{Y (m)}
For i= 1, 2, , K do
a
i
(m) from (2)
if speech is absent:
2
i
(m) from (1) end;
MMSE: si(m) from (5),
or ML: si(m) from (6)
end; end;

X(m) idct{[ s
1
(m),, , s
K
(m)]}
domain, where the ML approach results in a lower residual noise
level. From the spectrograms in Figure 3 it can be seen also that
the background noise is very efciently reduced, while the energy
of most of the speech components remained unchanged.
Figures 3 and 4 illustrate the results for a nonstationary, col-
ored lab noise. From our simulations and Figures 3, 3 and 4, we
conclude that the proposed methods perform very well for various
noise conditions such as for colored and/or nonstationary noises.
The estimation of a
i has an important impact on the perfor-
mance of the proposed SEAs. To illustrate this impact, we esti-
mate the Laplacian factor ai using the clean speech signal and call
this estimate as best value of ai. In Figure 5, noisy speech is
enhanced with these so-called best values. We will use the term
best value to refer to the SEA that processes the noisy speech
with these so-called best values, which theoretically provides the
best performance that can be achieved with this SE framework
under the Laplacian-Gaussian assumption. We can clearly see that
the residual noise level of this ideal case is much lower than the re-
sults from Figure 3. This illustrates the effectiveness of the SEA.
(a) Speech signal
0 2 4 6 8 10
1
0
1
(b) Noisy signal
0 2 4 6 8 10
1
0
1
(c) Enhanced signal (ML)
0 2 4 6 8 10
1
0
1
(d) Enhanced signal (MMSE)
0 2 4 6 8 10
1
0
1
(e) Enhanced signal using [6]
0 2 4 6 8 10
1
0
1
Time (sec)
(a) Speech signal
F
re
q
u
e
n
c
y
0 2 4 6 8
0
1000
2000
3000
4000
5000
(b) Noisy signal
0 2 4 6 8
0
1000
2000
3000
4000
5000
(c) Enhanced signal (ML)
0 2 4 6 8
0
1000
2000
3000
4000
5000
(d) Enhanced signal (MMSE)
0 2 4 6 8
0
1000
2000
3000
4000
5000
(e) Enhanced signal using [6]
0 2 4 6 8
0
1000
2000
3000
4000
5000
Time (sec)
Fig. 3. Enhanced speech corrupted by white Gaussian noise
(SNR=5dB), and corresponding spectrograms.
(a) Speech signal
0 2 4 6 8 10
1
0
1
(b) Noisy signal
0 2 4 6 8 10
1
0
1
(c) Enhanced signal (ML)
0 2 4 6 8 10
1
0
1
(d) Enhanced signal (MMSE)
0 2 4 6 8 10
1
0
1
(e) Enhanced signal using [6]
0 2 4 6 8 10
1
0
1
Fig. 4. Enhanced speech corrupted by nonstationary colored lab
noise (SNR=5dB).
(a) Speech signal
0 2 4 6 8 10
1
0
1
(b) Noisy signal
0 2 4 6 8 10
1
0
1
Time (sec)
(c) Enhanced signal (ML)
0 2 4 6 8 10
1
0
1
(d) Enhanced signal (MMSE)
0 2 4 6 8 10
1
0
1
Fig. 5. SEA results with the best value, white noise, SNR=5dB.
I - 299

Table 3. Spectral Distortion between the clean signal and signals
enhanced using different SEAs for various noise conditions.
Noise Input Input Proposed SEA best value of ai SEA
Type SNR SD MMSE ML MMSE ML in [6]
White 0dB 5.85 5.67 5.78 5.41 5.68 5.70
Gaussian 5dB 4.84 4.58 4.67 4.27 4.73 4.59
Noise 10dB 3.78 3.53 3.57 3.40 3.75 3.53
Lab 0dB 5.98 5.58 5.52 5.01 5.53 5.52
Noise 5dB 4.91 4.58 4.54 4.11 4.51 4.49
10dB 3.81 3.54 3.55 3.24 3.53 3.49
Computer 0dB 4.88 4.12 4.34 4.96 4.59 4.29
Fan 5dB 3.73 3.26 3.53 3.50 3.77 3.47
Noise 10dB 2.71 2.56 2.82 2.81 3.02 2.77
Spectral Distortion: We use Spectral Distortion (SD) as a crite-
rion for the performance of SEAs. The SD in decibels (dB) be-
tween two signals x(t) and y(t) with length N is dened by
SD(x(t); y(t))=
1
4N
N
64

i=1
255

k=0
20 |log
10
|Xp(k)| log
10
|Yp(k)||.
where Xp(k) and Yp(k) are the kth FFT frequency components of
the pth frame of
x(t)
||x(t)||
and
y(t)
||y(t)||
, respectively. Signals are di-
vided into frames of length 64 samples without overlapping. After
padding 192 zeros into each frame, the 256-point FFT is calcu-
lated. Table 3 presents SDs where the clean speech signal is com-
pared with the noisy input signal, two proposed enhanced signals
and the enhanced signal using the algorithm in [6]. In white and
lab noise conditions, the SD values for all these approaches are
better than the noisy speech. The result of the best value MMSE
approach is slightly better than those of the others. Only for the
fan noise in a high SNR condition does the SD result seem to be
unexpected. The reason is that the PSD of the fan noise has an
strong peak at a low frequency that results in a strong SD around
this frequency.
The Output SNR of the enhanced signal (or the noisy signal) y(t)
is dened by SNR = 10 log
10

N
k=1
x
2
(k)

N
k=1
(y(k)x(k))
2
. Table 4 com-
pares the enhanced signals using different approaches versus the
input SNR. As expected, the SNR performance of the best value
MMSE is the best (highest) in all noise conditions. The SNR im-
provement in the MMSE approach for high SNRs is higher than
that of other approaches.
6. CONCLUSION
An SEA is developed based on a Laplacian distribution for speech
and a Gaussian distribution for additive noise signals. The en-
hancement is performed in a decorrelated domain. Each compo-
nent is estimated from the corresponding noisy speech component
by applying a non-linear memoryless lter. The speech signal is
decomposed into uncorrelated components by the DCT or by the
adaptive KLT. It is assumd that the speech is stationary within 20-
40msec and the noise is stationary over a longer period of about
0.5sec. The proposed SEAs are based on the MMSE and the ML
approaches, respectively. The speech is then synthesized by the
IDCT or IKLT. Overall, proposed SEAs effectively reduce the ad-
ditive noise. At the same time, the proposed SEAs produce a lower
level of distortion in the enhanced speech when compared with the
Table 4. Comparison of SNR (in dB) of enhanced signals for var-
ious noise conditions.
Noise Input Proposed SEA best value SEA SEA
Type SNR MMSE ML MMSE ML in [6]
White 0dB 4.30 4.87 6.05 5.38 5.22
Gaussian 5dB 8.26 8.51 9.24 8.67 8.72
Noise 10dB 12.47 12.47 12.87 12.32 12.57
Lab 0dB 4.26 5.10 6.49 5.73 5.40
Noise 5dB 8.27 8.64 9.45 8.64 8.78
10dB 12.33 12.32 12.84 11.99 12.37
Computer 0dB 5.74 5.90 6.48 6.15 6.04
Fan 5dB 9.87 9.82 10.26 9.84 9.94
Noise 10dB 13.46 13.29 13.57 13.16 13.37
method in [6] that uses a complex Adaptive KLT. The comparison
of results with the method in [6] shows that the proposed SEAs
provide a better (or similar) performance. The performance cri-
teria of the proposed SEAs give similar results. The fact that the
SEAs with best value outperformed all the others, indicates that
the new proposed framework for SE could be further improved.
The computational complexity of the proposed SEAs is very
low compared with the existing algorithms because of the use of
fast DCT. In fact, most of the computationally complex parts are
the DCT and IDCT (the computational complexity the DCT and
IDCT is of the order of K log
2
(K), where K is the size of the
vectors). All our simulations and listening evaluations conrm that
the proposed methods are very useful for SE.
7. REFERENCES
[1] Y. Ephraim and D. Malah, Speech Enhancement Using a
Minimum Mean-Square Error Short-Time Spectral Amplitude
Estimator, IEEE Trans. Acoustics, Speech and Signal Pro-
cessing, vol. 32, no. 6, pp. 1109-1121, Dec. 1984.
[2] S. M. McOlash, R. J. Niederjohn, and J. A. Heinen, A Spec-
tral Subtraction Method for the Enhancement of Speech Cor-
rupted by Nonwhite, Nonstationary Noise, Proc. of IEEE Int.
Conf. on Industrial Electronics, Control, and Instrumentation,
vol. 2, pp. 872-877, 1995.
[3] I. Y. Soon and S. N. Koh, Low Distortion Speech Enhance-
ment, IEE Proceedings-Vision, Image and Speech Process-
ing, vol. 147, no. 3, pp. 247-253, June 2000.
[4] Z. Goh, K.-C. Tan, and B. T. G. Tan, Kalman-Filtering
Speech Enhancement Method Based on a Voiced-unvoiced
Speech Model, IEEE Transactions on Speech and Audio Pro-
cessing, vol. 7, no. 5 , pp. 510-524, Sept. 1999.
[5] U. Mittal and N. Phamdo, Signal/Noise KLT Based Ap-
proach for Enhancing Speech Degraded by Colored Noise,
IEEE Trans. Speech and Audio Processing, vol. 8, no. 2 , pp.
159-167, March 2000.
[6] A. Rezayee and S. Gazor, An Adaptive KLT Approach for
Speech Enhancement, IEEE Trans. Speech and Audio Pro-
cessing, vol. 9, no. 2, pp. 87-95, Feb. 2001.
[7] Y. Ephraim and H. L. Van Trees, A Signal Subspace Ap-
proach for Speech Enhancement, IEEE Trans. Speech and
Audio Processing, vol. 3, no. 4, pp. 251-266, July, 1995.
[8] S. Gazor and W. Zhang, Speech Probability Distribution,
IEEE, Signal Processing Letters, vol. 10, no. 7, pp. 204-207,
July 2003.
[9] S. Gazor and W. Zhang A soft voice activity detector based
on a laplacian-gaussian model, IEEE Transactions on Speech
and Audio Processing, vol. 11, no. 5, pp. 498-505, Sept. 2003.
I - 300