A Design of Speech Recognition System for Chinese Proper Nouns

• Main Authors: Bo-wun Cheng, 鄭博文
• Other Authors: Chih-Chien Chen
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/g5dg98

碩士 === 國立中山大學 === 電機工程學系研究所 === 106 === In today’s technological era, “Speech Web Searching” for person, event, time, place and object is a common necessity for our daily life. Although many searching websites can offer this service free of charge, it is a de facto practice to search “on line” due to their commercial profit models. Hence it is our hope to design a proper noun speech searching system that can be used “offline” to promote personal convenience. In this thesis, the mel frequency cepstral coefficients and linear predictive cepstral coefficients are used to extract the speech characteristics. Firstly, a training database of 2,699 two-syllable words are recorded. All the syllables are then used to design a syllable speech classifier. Secondly, in order to extract more detailed information of a syllable, the initial consonant and the final vowel for the syllable are applied to construct the two auxiliary classifiers. Finally, incorporating the phonotactic rules, the best candidate word is selected by using both the syllable and auxiliary classifiers. Furthermore, a multi-dimensional syllable tag is established for each noun. This technique can improve the missing or mistaken syllable error when the speech input is cross examined by the hidden Markov model, the phonotactics and the syllable tag. A Chinese proper noun database of approximate 950 thousand entries is collected for system evaluation. Under the Intel Core ™ i7-4712MQ notebook with 2.3 GHz CPU and the Windows 8.1 operating system environment, the speech system can obtain a 95.89% correct recognition rate.