Digital Signal Processing Reference
In-Depth Information
Normalized least mean squares (NLMS) algorithm is implemented for adaptive
noise canceller [
5
,
6
]:
Uk
ðÞ
;
t
Y
o
k
ðÞ
;
t
G
B
k
Þ¼G
B
k
ð
;
t
þ
1
ðÞþm
;
t
ðÞ
;
(12.6)
P
est
k
;
t
in which the time-frequency index returns to describe the update in short-time
Fourier transform domain. In (
12.6
), the adaptation term is controlled by the power
estimate of the input sensor signals:
X
2
2
P
est
k
ðÞ¼a
;
t
Pk
ð
;
t
Þþ
ð
a
Þ
Z
jj
;
1
1
(12.7)
i¼
1
where
a
is a forgetting factor. Then, the resulting system output is given by
ðÞG
B
k
Yk
ðÞ¼
;
l
Y
C
k
;
l
ðÞ
;
l
Uk
ð :
;
l
(12.8)
A high computational burden occurs in the TFGSC when the number of adaptive
filter coefficients is large enough to cover the signal path in a reverberant chamber.
A save/add method is applied to perform a linear convolution using FFT. It
necessitates a computationally efficient adaptive noise suppression filter while
keeping the advantage of TFGSC.
12.2.2 Perceptually Adaptive Noise Suppressor (PANS)
Based on TFGSC
The structure of the PANS based on TFGSC is shown in Fig.
12.2
. The PANS is
composed of three blocks: a fixed beamformer (FBF), a blocking matrix (BM),
and a perceptually adaptive noise suppressor (PANS). It is used to estimate the
spectral envelope (SE) of the desired speech signal. As shown in Fig.
12.3
,an
auditory filterbank such as the Mel-filterbank or the equivalent rectangular bandwidth
characterizes the PANS [
7
,
8
]. The filterbank is composed of band-pass filters
imaging the effect of auditory masking. Accordingly, specific frequency resolution
of a human auditory system is provided. As shown in Fig.
12.2
, the filterbank outputs
the auditory SE of the primary signal
Y
and that of the reference noise
U
.Then,an
adaptive filter estimates the noise SE
N
of the primary signal with the input
U
:
Given
the estimate
N
;
the spectral modification is executed to obtain the desired speech as
h
i
0
:
5
S
ax
¼
F
itp
Y
;
1
(12.9)
Search WWH ::
Custom Search