Robust voice activity detection using cepstral features

This paper reviews algorithms which rely on the analysis of time domain samples to provide energy and zero-crossing rates, together with more recent algorithms that use different methods for speech detection. We then examine a different approach using cepstral analysis, showing a high degree of amplitude and noise level independence. We show that a cepstral based algorithm exhibits a high degree of independence to levels of background noise and successful speech end-pointing can be achieved via thresholding cepstral distance measures. Through the use of a noise code-book we are able to provide a successful reference for Euclidean distance measures in the voice detection algorithm.<>