Real-time digital signal processing: implementations, ... changes in the input signal is limited by its internal clock rate, so that it may be slow to
Document source : notes.ump.edu.my
and y
x
k is the phase of measured noisy signal. It is sufficient to use the noisy speech
phase for practical purposes. Therefore we reconstructed the processed signal using the
estimate of short-term speech magnitude spectrum j
^
Skj and the phase of degraded
speech, y
x
k.
Substituting Equations (9.6.3) and (9.6.5) into Equation (9.6.4), the estimator can be
expressed as
^
Sk jXkj À EjVkj
Xk
jXkj
HkXk,
9:6:6
where
Hk 1 À
EjVkj
jXkj
:
9:6:7
Note that the spectral subtraction algorithm given in Equations (9.6.6) and (9.6.7)
avoids computation of the phase y
x
k, which is too complicated to implement in a
fixed-point hardware.
9.6.3 Implementation Considerations
A number of modifications are developed to reduce the auditory effect of spectral error.
These methods are spectral magnitude averaging, half-wave rectification, and residual
noise reduction. A detailed diagram for spectral subtraction algorithm is illustrated in
Figure 9.21.
Spectral magnitude averaging
Since the spectral error is proportional to the difference between the noise spectrum and
its mean, local averaging of the magnitude spectral
jXkj
1
M
X
M
i1
jX
i
kj
9:6:8
Noisy speech
x(n)
FFT
IFFT
Phase
Magnitude
averaging
Subtract
noise
Residual noise
reduction
Half-wave
rectification
Processed speech
^
s(n)
Figure 9.21 Detailed diagram of spectral subtraction algorithm
SPEECH ENHANCEMENT TECHNIQUES
433
Summary :
Spectral magnitude averaging Since the spectral error is proportional to the difference between the noise spectrum and its mean, local averaging of the magnitude spectral jXkj 1 M X M i1 jX i kj 9:6:8 Noisy speech x(n) FFT IFFT Phase Magnitude averaging Subtract noise Residual noise reduction Half-wave rectification Processed speech ^ s(n) Figure 9.21 Detailed diagram of spectral subtraction algorithm SPEECH ENHANCEMENT TECHNIQUES 433
Tags :
spectral,speech,phase,magnitude,aeraging,jxkj,noise,algorithm,noisy,subtraction,reduction,966,signal