REDUCTION OF DIFFUSE NOISE IN MOBILE AND VEHICULAR APPLICATIONS - DSP for In-Vehicle and Mobile Systems - page 159

and are Wiener filter parameters that have been optimized for
this application [4]. Also, is the uncorrelated noise spectrum

estimated from during speech pauses.

Figure 10-3. Block diagram of the subband processing block.

2.4 Voice Activity Detector

A Voice Activity Detector (VAD) has been employed to detect the noise-
only portions of the primary (noisy) input. It is a modified version of the
ETSI AMR-2 VAD [12] that has been implemented on the oversampled
WOLA filterbank [13]. As depicted in Figure 10-4, the WOLA filterbank
analysis results for the primary input are first grouped as a number of
channels and the energies of the channels m is the
frame index) are estimated. Given an estimate of the background noise
channel SNR is estimated. A
non-linear function maps the channel SNR to a voice metric V(m). Channel
SNR is also used to calculate a frame SNR and a long-term SNR. The voice
metric and the long-term SNR provide primary parameters for the VAD
decision. There is also a hangover mechanism in the VAD. A spectral
deviation estimator measures the deviation between the frame subband
energies and the long-term subband energies. When the deviation (averaged