The speech recognition technique for the whole Chinese vocabulary

A speech recognition technique for the entire Chinese vocabulary is discussed. Because of the transient nature of speech, a recursive algorithm is used instead of a framewise algorithm to analyze the speech signal. A novel recognition unit has been used as the broad phonetic representation to access the lexicon and to account for the effect of phonetic cues of consonants and vowels. Thirty-eight vowels and 21 consonants of Chinese were put into 13 and 6 clusters, respectively, according to some phonetic cues. The duration of consonants is also used to classify consonants broadly. For each consonant and vowel, a continuous hidden Markov model is designed. The model log-likelihood scores for consonant and vowel are used for the final recognition decision. The results show that the recognition technique presented is promising