Frequency-domain noise suppression also can be implemented in time-domain by
first decomposing the corrupted speech signal into a different frequency band using
bandpass filterbank. The noise power of each subband is then estimated during non-
speech periods. Noise suppression is achieved through the use of the attenuation factor
corresponding to the temporal signal power over estimated noise power ratio of each
subband. Since the spectral subtraction algorithm provides the basic concept of filter-
bank technique, it will be presented in detail in the next section.
9.6.2 Spectral Subtraction Techniques
Spectral subtraction offers a computationally efficient approach for reducing noise by
using the FFT. Assume that a speech signal s(n) has been degraded by an uncorrelated
additive noise v(n). As illustrated in Figure 9.20, this approach enhances speech by
subtracting the estimate of the noise spectrum from the noisy speech spectrum. The
noisy speech x(n) is segmented and windowed. The FFT of each data window is taken
and the magnitude spectrum is computed. A VAD is used to detect the speech and non-
speech activities of the input signal. If the speech frame is detected, the system will
perform the spectral subtraction and the enhanced speech signal
^
sn will be generated.
During the non-speech segment, the noise spectrum will be estimated and the data in the
buffer will be attenuated to reduce noise.
There are two methods for generating the output during non-speech periods: (1)
Attenuate the output by a fixed factor, and (2) set the output to 0. The experimental
results show that having some residual noise (comfort noise) during non-speech frame
will give higher speech quality. A possible reason for this is that noise present during
speech frames is partially masked by the speech. Its perceived magnitude should be
Data segmenting
and buffering
x(n)
Spectral
subtraction
Noise spectrum
estimation
VAD
Speech
Non-speech
Attenuation
^
s(n)
Figure 9.20 Block diagram of the spectral subtraction algorithm
SPEECH ENHANCEMENT TECHNIQUES
431