SPEECH RECOGNITION DEVICE AND COMPUTER PROGRAM

摘要

PROBLEM TO BE SOLVED: To recognize a speech of a utterance part of a speaker with high reliability in an environment wherein the speaker and another person have a conversation. SOLUTION: The speech recognition device 40 includes a sound model storage unit 64; storage units 66 and 68 which store bigram and trigram language models generated based upon a text DB 44 corresponding to a previously assumed utterance; a speech recognizing unit 42 which performs speech recognition by a statistical method using sound models and language models and outputs an N-best hypothesis; a GWPP computation processing unit 76 which computes a generalized word follow-up probability (GWPP); a word deleting unit 78 which deletes a word whose GWPP computed by the GWPP computation processing unit 76 is less than a threshold; and a re-scoring unit 82 which rescores respective hypotheses after words are deleted based upon generalized word follow-up probabilities of words included in the hypotheses and outputs a predetermined number of high-order hypotheses as speech recognition results. COPYRIGHT: (C)2008,JPO&INPIT