Mandarin and Engliah Speech Recognition System Using a Discriminative Bayesian Network

• Main Authors: Lai, Lance, 賴育昇
• Other Authors: Chung-Hsien Wu
• Format: Others
• Language: zh-TW
• Published: 1996
• Online Access: http://ndltd.ncl.edu.tw/handle/36681081158417674805

碩士 === 國立成功大學 === 資訊工程學系研究所 === 84 === In this thesis, a continuous word-based Mandarin and English recognitionsy stem based on a discriminative Bayesian network(DBN) is proposed. This system is ableto recognize Mandarin and English speech simultaneously. We collect four kinds of training databases, a Chinese 176-monosyllabledatabase, a Chines e balanced word database, and two English balanced worddatabases. Then each ut terance in these databases was segmented into phonemeunits for system training . In the training phase, a discriminative Bayesian network was used to mode leach phoneme. In the recognition phase, the one-stage algorithm was adopted f orrecognition, in which external transition was constrained by a grammar tree. Thus not only the recognition rate but also the response time were improved. In our experiments, we chose 100 Chinese and 100 English vocabularies as the outside testing database. The experimental results show that the top 1, top 3, and top 10 recognition rates achieved 90.6%, 91.6%, and 93% respectively.