連續中文字的語音辨識系統

Abstract

本論文描述一個語者獨立,連續中文語音辨識系統的實作.整個系 統分成兩個部份.第世個部份的功能是將中文句子切成獨立的中文單字.第 二個部份則將每個獨立單字辨識出來.在糸統中,切割單字是根據句子的頻 譜資訊及線性預估係數,而單字辨識方式是一種反傳遞網路模型的應用.而 且為了要解決大部份類神經網路語音辨識器因固定結構所產生的時間安排 問題,我們採用了最小扭曲切割方法.系統在小字彙已有不錯的結果.因此 未來要加強的是增加字彙. The thesis addressed an implementation of speaker independent, continuous Mandarin speech recognition system. The overall system consists of two seperate but interrelated parts. The function of the first part of the system is to segment the Mandarin sentence into individual Mandarin words; the second part of the system then recognize each of the individual words based on the results of the segmentation. In the system, the segmentation is based on spectral analysis of the sentence as well as LPC coefficients while the recognition is an application of Counter Propagation Network(CPN) model. Specifically, in order to solve the time-alignment problem in the fixed networkstructure of most ANN based speech recognizer, a minimal distortion segmentation(MDS) method is applied. Satisfactory results have been obtained for small vocabulary and future work is to consider a large size of vocabulary.