Digital Signal Processing Reference
In-Depth Information
where
p(H
0
Y
k
)
denotes the probability of speech absence given
Y
k
.Sincethe
expected speech spectrum under speech absence is zero, i.e.
E(X
k
|
|
Y
k
,H
0
)
=
0,
equation (11.42) can be simplified to,
X
k
=
E(X
k
|
Y
k
,H
1
)p(H
1
|
Y
k
)
(11.43)
E(X
k
|
Y
k
)
can be computed by a conventional spectral esti-
mator and equation (11.39), respectively.
Y
k
,H
1
)
and
p(H
1
|
11.2.6 Comparisons
Objective speech qualities for voice-active regions are evaluated in terms
of both segmental SNR (SEGSNR) improvement and Itakura-Saito distor-
tion (ISD). The SEGSNR improvement indicates the difference between the
SEGSNRs of the enhanced speech and the noisy input signals, in which the
SEGSNR is defined by,
#
$
&
'
M
−
1
(m
+
1
)N
−
1
x
2
(n)
(x(n)
10
M
!
SEGSNR
(
dB
)
=
log
10
(11.44)
− ˆ
x(n))
2
%
m
=
0
n
=
mN
where
N
and
M
aretheframesizeandthetotalnumberofframes,respectively.
The ISD is defined as,
a
x
R
x
a
x
ˆ
ISD
(
dB
)
=
10 log
10
(11.45)
a
T
ˆ
x
R
x
a
ˆ
x
ˆ
where a
x
and a
x
are the LPC coefficients of the desired and estimated speech
signals, respectively, and R
ˆ
is the autocorrelation matrix of the estimated
x
ˆ
signal.
For comparison, speech material of 64 seconds, mixed with vehicle and
helicopter noises of 0, 5 and 10 dB SNR were used. Enhancement processing
was applied every 10ms in the frequency domain by the five types of spectral
estimator: PSS, GBSS, ML, WF, and MMSE-LSA. The MMSE-LSA is further
classified, depending on the adoption of the speech presence uncertainty,
into MMSE-LSA-HD and MMSE-LSA-SD in which HD and SD denote the
hard and soft decision methods, respectively. The reference (the best possible
processed signal) is obtained using the original spectral amplitudes with the
phases of the noisy signal, because the ideal speech enhancement is achieved
with the original speech spectral amplitudes and the phases of the noisy input
speech.
The SEGSNR improvement and ISD for the vehicle and the helicopter noisy
signals are shown in Figures 11.2, 11.3, 11.4, and 11.5. From the analysis, it is
Search WWH ::
Custom Search