Endpoint detection of noisy speech based on cepstrum

A major cause of errors in automatic speech recognition (ASR) systems is the inaccurate detection of the beginning and ending boundaries of test and reference patterns. Accurate determination of endpoints of speech is not very difficult if the SNR is high. Unfortunately, most practical ASR systems must work with a small SNR, and the conventional speech detection methods based on some simple features, such as energy cannot work well in noisy environments. In this paper, cepstrum is used as the feature to detect the voice activity. Two algorithms for endpoint detection of noisy speech signal are proposed. The first one takes the cepstral distance as the decision thresholds instead of short-time energy. The second approach modified the HMM-based speech detector to make it adaptive to the change of noise. The experiments show that high accurate rates can be obtained.