A vowel-driven connected Mandarin digit recognition system

A vowel-driven connected Mandarin digit recognition system is presented in this paper. Since Mandarin is a monosyllabic language, each syllable has a consonant (which may be null) and a vowel. Hence, a concatenated digit string is composed of a series of consonants and vowels. A segmentation algorithm is proposed to segment a connected digit string into its consonant and stable-region parts, where a stable region may contain more than one vowel. The stable-region parts are recognized first. Mandarin digits, except (1, 7) and (6, 9), can be recognized just by their vowel parts. The digit pairs (1, 7) and (6, 9), which have the same vowel parts, are further distinguished by the special acoustic characteristic of the consonant part. The recognition system has been tested in a speakerdependent mode under two different cases; known string length and unknown string length. String error rates of 5.3% and 7.4% were achieved in the knownlength and unknown-length cases, respectively.