摘要 |
PROBLEM TO BE SOLVED: To perform large vocabulary voice recognition with high performance and at a high speed by utilizing a phoneme model depending on a phoneme environment and a phoneme duration length model based on the view point of an A* retrieval method and reducing a calculation amount. SOLUTION: This voice recognition system is constituted of a heuristics calculation part 8 for turning the information of the phoneme array structure of Japanese to a language restriction and performing score calculation from the start end of voice to a terminating end in time synchronism with input voice and a candidate search part 10 for turning the phoneme and information of a recognition object vocabulary to the language restriction and searching candidates from the voice terminating end to the start end while considering heuristics 9 obtained in the heuristics calculation part 8. At the time of recognizing the voice, as the respective language restrictions, a finite state automation whose symbol is the state number of HMM depending on the phoneme environment, obtd. by sharing an entirely or partially common state number sequence and expressing in a network shape is used. |