A Vocabulary-Size Independent Mandarin Word Recognition System

• Main Authors: Eric Jan, 詹曄暐
• Other Authors: Jhing-Fa Wang
• Format: Others
• Language: zh-TW
• Published: 1993
• Online Access: http://ndltd.ncl.edu.tw/handle/85252080954914415404

碩士 === 國立成功大學 === 資訊工程研究所 === 81 === The difficulties in large vocabulary size speech recognition system are the high computation load and low recognition rate. In this thesis,a vocabulary-size independent recognition system is proposed to overcome the above two problems.The recognition unit of conventional word recognizer is usually "word",but we choose "phoneme" as recognition unit. In other word,each syllable is divided into two parts - consonant and vowel. In this system,several schemes are introduced.First,we propose a fast Mandarin syllable segmentation algorithm to label the consonant and vowel parts of each syllable. Second,a vowel preprocessor is constructed to accelarate recognition speed of vowel.Third,the Bayesian reference template is constructed with variable gaussian nodes and transition probabilities to avoid unreasonable transitions and to effectively decrease error rate. Finally,we establish an efficient lexicon for word training and word matching. Experimental results show the effectiveness of the proposed system.The average word recognition rates are 70％ and 83.4％ for top 1 and top 5 respectively.The average recognition time for each syllable is about 0.61 second.