Robust Endpoint Detection in Mandarin Based on MFCC and Short-Time Correlation Coefficient

Endpoint detection is the process of cutting the signal into pieces according to its syllables or identifying the important part of a speech segment for further processing, which is a significant part in speech recognition. Recognition accuracy and robustness of the traditional methods for example, the short-time energy spectral and Zero-pass Ratio analysis will decrease sharply with the Signal-to-Noise-Ratio (SNR) going down. Base on some special feathers of Mandarin, a new robust algorithm with the combined using of MFCC and Short-time Correlation Coefficient Analysis will be described in this paper, which has excellent noise immunity. The experimental results show that even with a lower SNR, the recognition accuracy is still higher and more stable than the other algorithm, and have great potential in speech signal processing.