Voice Activity Detection - Digital Speech: Coding for Low Bit Rate Communication Systems - page 365

The a posteriori SNR γ k fluctuates highly from frame to frame because of
the high fluctuation of the short-time spectral amplitude

.Ontheother
hand, the apriori SNR ξ k changes slowly due to the smoothing effect. As
the value of α increases, ξ k becomes smoother. The variations of γ k and
ξ k balance each other in the calculation of k and, consequently, result in
enhanced performance for the VAD. The DD estimator for the apriori SNR
is therefore useful not only for avoiding the musical noise phenomenon in
speech enhancement [15], but also for reducing the error rate in voice activity
detection.

| Y k |

10.3.1 AnalysisandImprovementof theLikelihoodRatioMethod
The behaviour of the LR in equation (10.3) with respect to the apriori and
a posteriori SNRs, is shown in Figure 10.14. The ML estimator [12] results
in lower performance in comparison with the DD estimator because of the
inherent high-fluctuation of the a posteriori SNR. The LR employing the DD
estimator has the following properties:

• If the a posteriori SNR is very high, i.e. γ k 1, and the range of the apriori

SNR is limited, the LR becomes very high, i.e. k

1.

•

If the a posteriori SNR is low, i.e. γ k < 1, the apriori SNR becomes a key
parameter in the calculation of the LR.

35

30

25

20

15

10

5

0

− 5

− 10

− 15

− 15 − 10 − 5 0 5 10 15

A priori SNR (dB)

Figure 10.14 Likelihood ratio vs aprioriSNR vs aposteriori SNR (the solid lines from
the top represent aposteriori SNRs of 15, 10, 5, 0, −5, −10, and −15 dB, respectively)