Speech Enhancement - Digital Speech: Coding for Low Bit Rate Communication Systems - page 386

found that WF- and MMSE-based methods give better results than the other
methods.
For the noisy input signal in Figure 11.6, the spectrograms of different
enhancement methods, showing the characteristics of the residual noise, are
shown in Figures 11.9, 11.10, 11.11, 11.12, 11.13, and 11.14. The spectrograms
of the noise-free and reference signals are shown in Figures 11.7 and 11.8,
respectively. For the PSS and ML-based methods, severe musical noise gives
irregular spots in the spectrograms in Figures 11.9 and 11.11, respectively.
The GBSS method with ν 0 . 1 reduces the musical tones to

a moderate level (see Figure 11.10), compared with the PSS and ML methods.
The WF-based method gives a further reduction in the level of the residual
noise as shown in Figure 11.12. Using the MMSE-STSA-based method, it
is possible to further eliminate the musical noises (see Figure 11.13). Even
though the level of the overall residual noise of the MMSE-STSA is slightly
higher than that of the WF method, the sound quality of MMSE-STSA is
perceptually more comfortable than that of the WF method. The higher
speech quality is due to further reduction in tonal signals. Combining the
soft-decision technique with the MMSE-based method, it is possible to reduce
the overall level of the residual noise as shown in Figure 11.14.

= 2, α = 4, and β =

11.2.7 Discussion
Ephraim andMalah's speech enhancement method gives higher performance
mainly due to the DD-based apriori SNR estimation. Cappe [16] has shown
its usefulness for eliminating musical noise phenomena through behavioural
analysis. From interpretation of equation (11.6), it is not difficult to see
that ξ k is a smoothed version of γ k .The a posteriori SNR γ k shows high
fluctuation from frame to frame, while ξ k changes slowly. By exploiting the
characteristics of the two SNRs, γ k and ξ k , improved performance in speech
quality is achieved.
The WF produces better performance than either GBSS- or ML-based
methods. The reason behind this better performance is also due to the
DD-based apriori SNR estimation used in the gain function of the WF. The
usefulness of the DD-based apriori SNR can also be applied to a posteriori SNR-
based speech enhancement methods, such as GBSS- and ML-based spectral
estimators, by replacing the a posteriori SNR with the apriori SNR [17] as,

= ξ k γ k + 1 (11.46)

Although substantial reduction of musical noise is achieved by the WF-based
method, it is observed that the musical noise is not completely removed (see
Figure 11.12). It is also possible to show that the musical noise phenomenon
exists in the apriori SNR-based speech enhancement using equation (11.46).