Effects of Segmentation of Speech on Speaker Recognition

• Main Authors: Yu Hsiang-Yun, 余祥雲
• Other Authors: Hsu Chien-Min
• Format: Others
• Language: zh-TW
• Published: 1999
• Online Access: http://ndltd.ncl.edu.tw/handle/u2hc6k

碩士 === 中央警察大學 === 刑事警察研究所 === 87 === This thesis presents an experimental evaluation of the speaker recognition by examine various types of continuous mandarin speech. The terms of speech features are based on the cepstral coefficients, which are derived from LPC (Liner Predictive Code) parameter. As the result, regard the overlapping continuous words as a comparative unit will improve the ability of recognizing whether the speeches are from the same speakers or not. In addition, it will be helpful for developing automatic segmenting speech systems. According to the experiment, it also present a standard how to recognize speakers：If the comparative unit is the type of overlapping continuous words, we can make an identification conclusion by the distortion is under 6, make an elimination conclusion by the distortion is over 16, and make an inconclusive conclusion by the distortion is between 6 and 16. If the comparative unit is not the type of overlapping continuous words, we can make an identification conclusion by the distortion is under 2, make a probable identification conclusion by the distortion is between 2 and 6, make a probable elimination conclusion by the distortion is over 16, and make an inconclusive conclusion by the distortion is between 6 and 16. Results indicate that this is a very promising method for automatic speech recognition of speaker system.