Noisy Speech Signal Endpoint Detection And Phonological Separation

This paper includes three important speech signal technologies: endpoint detecting, pitch detecting and initial/final segmentation.Endpoint Detecting: Pronunciation of speech signal can be separated from Noise by endpoint detecting. This paper has researched and solved two problems of speech signal. One is "DC-offset of Speech signal 'Zero' level" and the other is "Wrongly detecting the endpoint of words" . The endpoint detecting arithmetic can be also used in the segmentation of multi-word. Short-time logarithm energy with a constant has be used in the place of short-time energy of classical arithmetic in this article. The quotient and product of short-time energy and short-time zcr(zero crossing rate) are basic parameters of the arithmetic.Pitch Detecting: Pitch is one of most important parameters of speech signal. A new function(ACF/AMDF) arises from the division of ACF(Auto Correlate Function) and AMDF(Average Magnitude Difference Function).The new function ACF/AMDF is the kernel function in the pitch detecting arithmetic. Pitches of speech signal are detected by ACF ,AMDF and ACF/AMDF. Then Final results are got. Robustness of endpoint detecting and pitch detecting arithmetic is enhanced after taking these measures in this paper.Initial/final segmentation: Initial/final segmentation is important work of Chinese speech signal. This paper separates initial and final of Chinese speech following initial without pitch and final with pitch. Differing from others, the arithmetic in this paper has no doors so adaptability is improved.The mended arithmetic can be applied to speech recognizing, speaker recognizing and so on.