Research on voice activity detection in burst and partial duration noisy environment

Voice activity detection aims at detecting speech in noisy environment and is very important for speech recognition. In this paper, two novel methods are proposed for finding out burst and partial duration noisy signals in order to detect real speech from non-stationary noise and to improve the performance of continuous Mandarin speech recognition system. For the burst and low-energy noise adjacent to speech segment, a method using initial and final part of Chinese syllable is applied to detect accurate endpoints of speech based on traditional short-time energy and zero-crossing rates. For the isolated noise with relatively high energy, a short-time energy zero product based method is used. Both of the methods use time-domain features and have low computational complexity. Experimental results show that the system using proposed methods has improved accuracy in voice activity detection.