Digital Signal Processing Reference
In-Depth Information
Table 10.2
Comparison of speech and silence detection error rates of SLR-based,
LR-based, AMR2, and G.729B VADs
Detection error rate (%)
Vehiclenoise
Babblenoise
SNR
(dB)
VAD
Inactive Onset Offset
SAS
Inactive Onset Offset
SAS
5
SLR
13.87
6.42
7.51
0.00
29.40
2.43
4.93
0.52
LR
4.49
12.88
30.92
0.00
46.25
6.90
27.77
2.49
HO
∗
LR
+
5.33
12.05
12.86
0.00
46.50
4.52
11.28
1.48
AMR2
18.64
9.13
0.00
0.00
41.66
4.75
0.26
0.00
G.729B
8.58
70.23
60.21
5.14
48.17
56.79
45.88
5.12
15
SLR
17.12
3.48
0.73
0.00
29.20
2.05
0.00
0.00
LR
5.07
5.34
19.85
0.00
41.76
3.70
16.83
0.08
LR
+
HO
7.52
4.75
6.80
0.00
42.67
3.32
4.18
0.00
AMR2
20.15
3.78
0.00
0.00
51.53
2.19
0.26
0.00
G.729B
8.57
31.19
39.41
0.00
49.79
25.90
32.73
0.00
25
SLR
23.01
2.82
0.00
0.00
30.77
1.54
0.00
0.00
LR
6.64
3.29
11.79
0.00
34.38
1.54
8.75
0.00
LR
+
HO
10.94
1.56
2.75
0.00
36.45
0.89
1.59
0.00
AMR2
20.28
2.68
0.00
0.00
20.61
2.31
0.12
0.00
G.729B
8.85
12.75
19.06
0.00
44.30
11.34
15.49
0.00
LR
+
HO means LR-based VAD with the hangover scheme
10.4 Summary
In this chapter, standard VAD techniques as well as LR- and SLR-based VAD
have been reviewed. Through performance evaluation of the standard VAD
methods, including G.729B, GSM-EFR, AMR1, AMR2, and IS-127, it has been
shown that both AMR1 and AMR2 produce relatively high and consistent
performances over various noise sources and levels. On the other hand,
statistical-model-based LR VAD, performs well but may have a problem at
the offset regions of speech signals which may be solved with a hangover
in the decision making. The SLR method newly-proposed by Cho [16, 17]
has overcome this problem without the need for a hangover. SLR VAD has
comparable performance to AMR2.
Bibliography
[1] ITU-T (1996)
A silence compression scheme for G.729 optimised for terminals
conforming to ITU-T V.70
, ITU-T Rec. G.729 Annex B.
Search WWH ::
Custom Search