Speech Endpoint Detection Technology Research Under Low Signal To Noise Ratio

With the development of information technology,nowadays more and more hardidentified electronic devices appear in the society,in order to convert the voice signal into theright voice information,especially in noisy environment,including high noise and lownoise,we must have high identification methods,while we get the voice information wanted.However,the technology of speech endpoint detection is a key part of speech recognitiontechnology,which plays a very important role in speech colding,recognition,enhancement,enhancement,speech recognition,it directly affects the subsequent work.The existing methods of speech endpoint detection had good detection effects in highSNR environment,but as the noise increased,its detection performance have droped more andthe invalidation of some methods came, so the study of method of speech endpoint detectionwas of great significance in the low SNR environment.This paper introduced the recognition process and the endpoint detection of speechsignal, and then according to a series of features such as the time domain,frequency domainfeature,the characteristics of nonlinear and multi features,it introduced the common speechendpoint detection method and show the basic knowledge and the experimental and analysisof some simulation data so it provide a theoretical basis for detection method in low SNR,The endpoint detection algorithm perfectly combine energy and spectrum-zero rate short-termenergy and power spectral entropy and apply it to recognize the dial tone,such as bank cardnumbers, passwords and phone number. At the same time endpoint detection for recorded dialkeys by means of Matlab tools and spectral analysis,finally it get the recognition result. In thelow SNR outdoor circumstances, detection results of combining short-time energy zerocrossing rate with power spectral entropy was better than detection results of short-time energy zero crossing rate, the recognition rate reached more92%.The endpoint detection recognition for dial key, for its inaccurate problem in start andend point and the end point, or lower SNR identify, We have to improve on the conventionalmethod. I am currently researching and developing on the part.