Toward speech recognition of continuously spoken Chinese sentences.

Abstract

A undamental speech recognition procedure for continuously spoken simple Chinese sentences is described and the method of specific regions is proposed. Phonemes are identified every 10ms by extracting numbers of zero-crossings, PARCOR coefficients, F1 and F2 etc. from speech waves. In Chinese language, each Chinese character is pronounced as a monosyllable and has definite meaning. Using these characteristics, continuously spoken speech waves are divided into monosyllables, and each vowel segment of monosyllables is partitioned into8minor segments. The average first2formant frequencies of each first minor vowel segment point out a specific region on the F1-F2 plane. Since this region decides a group of monosyllables or Chinese characters which have similar vowels, a monosyllable can be identified from them. Moreover, use of a syntactic state transition network improves recognition scores of sentences. Average recognition scores of130charactersand33sentences uttered by3male adults are90.7% and75.7%, respectively.