Statistical Analysis of Mandarin Acoustic Units and Automatic Extraction of Phonetically Rich Sentences Based upon a Very Large Chinese Text Corpus

Abstract

Automatic speech recognition by computers can provide humans with the most convenient method to communicate with computers. Because the Chinese language is not alphabetic and input of Chinese characters into computers is very difficult, Mandarin speech recognition is very highly desired. Recently, high performance speech recognition systems have begun to emerge from research institutes. However, it is believed that an adequate speech database for training acoustic models and evaluating performance is certainly critical for successful deployment of such systems in realistic operating environments. Thus, designing a set of phonetically rich sentences to be used in efficiently training and evaluating a speech recognition system has become very important. This paper first presents statistical analysis of various Mandarin acoustic units based upon a very large Chinese text corpus collected from daily newspapers and then presents an algorithm to automatically extract phonetically rich sentences from the text corpus to be used in training and evaluating a Mandarin speech recognition system.