A new algorithm for voice activity detection

In this paper, a new VAD (voice activity detection) algorithm based on spectral estimation for the noise suppression system is presented, which divides an input speech signal into specific frequency bands. On the basis of the estimated signal and noise in each sub-band, a signal-to-noise ratio (SNR) which represents a sub-decision signal is calculated for each sub-band. The VAD decision is formed by utilizing these sub-decision signals. Moreover, the VAD decision during instances of sudden increase in background noise level is improved by continually monitoring the deviation of the spectral energy. Extensive experiments indicate that this newly proposed method works well.