Mohammad Hossein Moattar - Laboratory for Intelligent Signal and Speech Processing, Computer Engineering and IT Dept. Amirkabir University of Technology (AUT) Tehran, Iran
Mohammad Mehdi Homayounpour - Laboratory for Intelligent Signal and Speech Processing, Computer Engineering and IT Dept. Amirkabir University of Technology (AUT) Tehran, Iran

This paper presents a set of voice activity detection (V AD) methods, that are easy to implement, robust against noise, and appropriate for real-time applications. The common characteristic is the use of a voting paradigm in all the proposed methods. In these methods, the decision on the voice activity of a given frame is based on comparing the features obtained from that frame with some thresholds. In the first method, a set of three features, namely frame energy, spectral flatness, and the most dominant frequency component is applied. In the second approach however, the spectral pattern of the frames of vowel sounds is used. To use the strengths of each of the above methods, the combination of these two decision approaches is also put forth in this paper. The performance of the proposed approaches is evaluated on different speech datasets with different noise characteristics and SNR levels. The approaches are compared with some conventional V AD algorithm such as ITU G. ۷۲۹, AMR and AFE from different points of view. The evaluations show considerable performance improvement of the proposed approaches.

voice activity detection, spectral flatness, vowel spectral pattern, noise robustness, vowel sounds