A STUDY OF SPEAKER RECOGNITION

• Main Authors: Liu, Chi-Shi, 劉繼謚
• Other Authors: Wang, Hsiao-Chuan
• Format: Others
• Language: zh-TW
• Published: 1996
• Online Access: http://ndltd.ncl.edu.tw/handle/79630296959684505857

博士 === 國立清華大學 === 電機工程研究所 === 84 === This dissertation investigates the techniques in speaker recognition. They are the discriminative training algorithm, the scoring function and the acoustic segment based probabilistic model.Some new methods are derived for improving the accuracy rate. In discriminative training,we construct the hidden Markov model for each speaker.By taking into accountthe models of other competing speakers so that speaker separation is enhanced. The optimization solution can be obtained by using a probabilistic descent algorithm. In order to improve the performance of conventional scoring algorithm, we propose a new scoring method for speaker verification.This method is derived from the Bayes test for minimum risk to attain the objective of minimizing the error probability. The conventional frame based probabilistic models use the instantaneous spectral information of each individual frame. Instead, we proposed a new segment model which uses both instantaneous spectral information and previous spectral information. The fixed length segment model is a simple approach to get performance improvement. A more precise method is to consider the acoustic segments. In this method, speech signals are represented by the orthogonal polynomial function. An iterative algorithm is proposed to segment and model the speech signals. The segment boundaries can be automatically detected according to the characteristics of speech signals so that we can generate the segment based probabilistic models. A 100-speaker digit database collected through the public switching telephone network is used for a series of experiments to show the effectiveness of our proposed methods.