Chinese speech recognition system with very large vocabulary

This paper describes a speaker-dependent Chinese speech recognition system using the SUN-SPARC10 platform with a very large vocabulary. This system was based on the semi-continuous hidden Markov model (SCHMM). Considering the special characteristics of Chinese, this paper creatively proposes a new revised beam search algorithm, named vowel part lookahead beam search algorithm. An experiment proved that looking ahead for a vowel part in an utterance could give valuable information about the optimality of the current partial path, with this new information a more robust and harder pruning can be done