Abstrato

A voice activity detection algorithm based on spectral entropy analysis of sub-frequency band.

Zhang Yuxin, Ding Yan

This paper proposes an effective voice activity detection (VAD) algorithms in low SNR noise environment. Traditional short-term energy and zero-crossing rate can only get high performance at high SNR environment. The spectral entropy algorithm is used to detect stationary noise signal, which is based on the inherent steady characteristics of noise signal. The whole spectrum is divided into some sub-bands, and then, the entropy value of sub-bands are computed separately. Since the voice change is stronger in some frequency bands, the sub-frequency band is extracted for detecting endpoint. The experimental results show that proposed method greatly improves performance of VAD at low SNR environment.