BASED ON THE SUB-SYLLABLE FRAMEWORK FOR MANDARIN RECOGNITION

摘要

This paper proposed a new system design for Mandarin base-syllable recognition with two major features. Firstly, we made use of the sub-syllabic parts that form the Mandarin syllables, namely, the 21 INITIALS and 38 FINALs, as the basic speech units for recognition. Compared to the systems that used the set of 408 whole-syllables, the proposed system reduces the amount of speech units to less than 15%. Secondary, to develop recognition algorithms with higher efficiency, in addition to the commonly used statistical pattern recognition approach, we incorporate the concept of acoustic-phonetic approach to make full use of our knowledge of the Mandarin syllable structure and sound classification. Experimental results showed that the proposed system has advantages of improved accuracy rate and significant saving of computation time, memory space, and the amount of the required training data for speaker adaptation.