Research On Endpoint Detection For Mongolian Speech Based On Support Vector Machines

With the development of network and multi-media technology, speech technology is increasingly important in practical application and has unprecedented improvements in its content and depth. Speech recognition can be applied to many fields including industry, commerce, banking, services, military, medical as well as daily life. Endpoint detection is the first key technology in preprocessing stage of speech recognition system. Endpoint detection is to distinguish speech signal from non-speech signal in complex application environment so as to find the beginning and end of speech. Support Vector Machine (SVM) is a kind of new technology belongs to Data Mining. This paper applies SVM to endpoint detection of speech and overcomes some trivial problems and inaccuracy provoked by setting threshold.Mongolian is official language of Inner Mongolia Autonomous Region, and it is pervasive in daily life and cultural communication of Mongolian race. Therefore, research on Mongolian speech recognition plays the key role in Mongolian information processing. This paper analyzes the characteristics of Mongolian based on speech endpoint detection for Chinese and Tibetan, and proposes SVM to distinguish speech signal from non-speech signal in platform of MATLAB. Endpoint detection of Mongolian is realized and results show that it is effective for SVM to detect endpoint of speech. In addition, researches on capacity of short-time energy, short-time average zero crossing rate and Mel-Frequency Cepstrum Coefficient (MFCC) in distinguishing speech signal from non-speech signal are referred in this paper, which establish substantial foundations for subsequent work.