Intelligent Language Modeling and Processing in Mandarin Speech Recognition

• Main Authors: Yen-Ju Yang, 楊燕珠
• Other Authors: Lin-Shan Lee;Keh-Jiann Chen
• Format: Others
• Language: zh-TW
• Published: 1993
• Online Access: http://ndltd.ncl.edu.tw/handle/67737403331273368411

碩士 === 國立臺灣大學 === 資訊工程研究所 === 81 === The purpose of this thesis is to develop an intelligent language modeling and processing system in Mandarin speech recognition, so that to get high recognition rate, high speed of processing time and experience from tested sentence. In this thesis, the recognition method is based on bigram Markov model under various smoothing techniques and various methods to incorporate unigram Markov model. In order to improve recognition rate, speed and the reliability, the special feature of this thesis is to propose some improve methods : syllable filter, learning, dynamic training and cache. They can make the system more and more intelligent. A language modeling and processing system has been developed on SUN workstation. It has tested some news on newspaper, articles in magazine and short story. It has following advantages : . high correct rate The character recognition system has average recognition rate 87.83%. When the whole text has filled into cache, we test the text second time, a very high recognition rate 98.87% can be obtained under the cache effect. Then we test the text third time someday later, the average recognition rate is 92.12% under learning effect. The word recognition system with very large vocabulary dictionary has average recognition rate 91.31%, 99.71% and 94.27% (calculated by character unit). . high speed The character recognition system needs 0.002 second on average to recognize a character. While the word recognition system needs 0.03 second on averge to recognize a word. . flexible data structure The system can switch between different models easily.．module design　 The system can use different algorithm in every module easily.