Abstract
This paper presents a method for the recognition of lexical tones in Mandarin speech based on vector quantization and hidden Markov models. A Kay Visi-Pitch 6087DS is used to extract the fundamental frequency (F 0 ) contour. The features for the recognition of lexical tones are derived from an F 0 contour within the voiced part of a syllable. Markov models are generated for the four monosyllabic tones and for 15 disyllabic tone pairs based on these derived features. A Viterbi algorithm is applied to find the probability scores of a test tone scored by the models during the recognition phase. In speaker-independent tone recognition experiments, the average recognition rate was 97·9% for isolated monosyllabic words, 92·9% for disyllabic words, and 91·0% for trisyllabic words. The decline comes from the tone variation and the coarticulation between syllables. © 1989.