METHOD AND SYSTEM FOR SPEECH RECOGNITION, AND RECORD MEDIUM

摘要

PROBLEM TO BE SOLVED: To improve the response of recognition processing by starting recognition postprocessing before the end of speaking in a voiceless section is detected. SOLUTION: A matching part 6 matches a recognition candidate selectively generated by a candidate generation part 4 against a voice feature string found by a feature extraction part 2 to find probability for each recognition candidate. A pruning process part 7 deletes and discards recognition candidates having low probability from a candidate storage part 5 according to the matching result. A one-path candidate detection part 8 detects a recognition candidate which has only one path of a word string defining recognition algorithm among recognition candidates left in the candidate storage part 5 after the pruning process. A recognition postprocessing part 10 performs the recognition postprocessing on the basis of the recognition candidate having only one path as an origin at the point of time when the one-path candidate detection part 8 detects the candidate.