Order statistics filtering-based real-time voice activity detection algorithm

In this paper, we propose an effective real-time voice activity detection algorithm. It makes use of the subband spectral entropy as the speech/noise discrimination feature. The speech spectrum is divided into several subbands at first. Then, the spectral entropy of each subband is estimated. We apply order statistics filters (OSF) to a sequence of the subband entropies to obtain the spectral entropy of each frame. The speech/noise classification is based on the spectral entropy. The experimental results show that the proposed algorithm can distinguish speech from noise effectively and improve the performance of automatic speech recognition (ASR) system significantly. It is proved to be robust under various noisy environments and SNR conditions. Moreover, the proposed algorithm is of low computational complexity which is suitable for embedded ASR system application.