Voice Activity Detection - Digital Speech: Coding for Low Bit Rate Communication Systems - page 359

improvement in channel capacity by controlling the radio transmission power
to reduce co-channel interference. TIA/EIA released two kinds of RDA for IS-
96 and IS-127, called8 kb/sQualcommcode-excited linear prediction (QCELP)
and enhanced variable rate codec (EVRC), respectively. In the North Ameri-
can CDMA standard, IS-127 RDA supports three rates: 1, 1/2, and 1/8. Active
speech is encoded in 1 or 1/2 rate and background noise is encoded in 1/8
rate. The RDA of IS-733, called 13 kb/s QCELP, is the same as IS-127.
As input parameters, IS-127 RDA uses two subband energies with the
long-term prediction gain. Firstly, it calculates the smoothed subband energy
using a first-order AR-model. Subsequently, the signal and noise energies for
each subband are adapted depending on the long-term prediction gain. In
other words, the signal energy is actively adapted to the current input if the
prediction gain is relatively high. On the other hand, if the gain is relatively
low, it increases the noise adaptation rate. Using the two subband energies
of the signal and noise, each subband SNR is calculated. The final rate is
determined by comparing the SNRs with adaptive thresholds depending on
the level of background noise and the SNR of the previous frame, followed
by a hangover. The block diagram of IS-127 RDA is shown in Figure 10.7.

10.2.5 PerformanceComparisonofVADs
The five standard VAD algorithms have been evaluated in terms of detection
error rates for speech and silence. The test data was 96 seconds of speech,
filtered by the modified IRS, and then mixed with vehicle and babble noises
of 5, 10, 15, and 25 dB SNR. The active and inactive regions of the speech
material were marked manually. The proportions of the inactive and active
regions of the speech material were 0.43 and 0.57, respectively. The VAD
decision is carried out every 10ms in the cases of G.729B and AMR2, and
every 20ms in GSM-EFR, AMR1 and IS-127. With slight modification to
the AMR2 source code, it is possible to obtain 10ms results because AMR2
basically conducts the detection every 10ms and then returns 20ms results
using a logical combination of the two 10ms results. In handling the multiple

Thresholds

Signal
Energy
Update

Band
Energy
Smoothing

SNR
Computation

Threshold
Comparison

SNRs

Noise
Energy
Update

High / Low
Band Energies

Active/
Inactive/
Mixed

Figure 10.7 Block diagram of TIA/EIA IS-127 RDA