Multi-Decision Sub-Band Voice Activity Detection for Speech Enhancement
ALAN H. DAVIS, Siow Yong Low, Sven Erik Nordholm · 2008
This paper compares the performance of the recently de-veloped multi-decision sub-band voice activity detector (MDSVAD) [1], with the ITU G.729B voice activity detector (VAD) scheme [2] and the SNRVAD scheme [3] for a speech enhancement application. The study shows the importance of more detailed VAD decisions in the time-frequency plane to better maintain speech features. This is in keeping with the observation that typically a speech signal will not simul-taneously excite all frequency components at any one time instant. Here, the MDSVAD exploits the spectral structure of speech versus background noise to make independent voice activity decisions in separate subbands, resulting in multiple decisions for any frame. Results show that the decisions in separate sub-bands reduce the infamous musical tones sig-nificantly in the conventional spectral subtraction algorithm compared to the other two VADs. 1.