Voice Activity Detection with Focus on Low SNR and Transient Noise
Bendik Paulsrud · 2013
Voice activity detection is a technique to detect the presence or absence of human speech in a waveform. The purpose of this work has been to evaluate the performance of a voice activity detection (VAD) algorithm under low signal-to-noise ratio (SNR) and to imple- ment a transient noise detection (TND) method. The algorithm chosen to be evaluated was a algorithm described in [1], later referred to as GMM VAD. The GMM VAD was compared with ITUs VAD described in the G.729 standard [2]. The main focus of the work has been on the performance in a video conference environment with low SNR and with transient noises present. The algorithms were implemented and tested in MATLAB [3]. The performance of the algorithms were evaluated using two different speech databases; one large database named TIMIT [4], providing statistical significant results and one small database providing more realistic results for verification. Compared with the ITU VAD, the GMM VAD has in general a lower speech hit rate while maintaining a higher non-speech hit rate. A novel method to detect transient noises was designed in combination with the GMM VAD. The TND algorithm proved to deliver good results, with a transient hit rate over 90 % in most test situations. Most important, it showed the best results in situations when most needed. Further work consists of finding better ways to update the speech/non-speech model in the GMM VAD and improve the decision routine for TND. The transient noise feature shows promising results and a better, more dynamic decision routine would be beneficial.