Performance comparison of new endpoint detection method in noise environments

Endpoint detection is the most important step of speech recognition. A good endpoint detection method can not only increase the success rate of speech recognition but also save the data storage space and reduce data processing time. This paper tests a new endpoint detection method based on linear prediction coefficient and makes performance comparison with methods based on short-time energy, short-time cross-zero and short-time autocorrelation in noise environments (including laboratory environment, server side environment, inner vehicle environment at idle speed). The result shows that new endpoint detection method based on linear prediction coefficient has good robustness on the starting point detection of speech signal.