| 摘要 |
PROBLEM TO BE SOLVED: To find reliability of a recognition result of a long section of each utterance, each sentence, etc., rather than a short section of a word, a syllable, etc. SOLUTION: In a speech recognition reliability estimating device, a speech recognition section 6 divides an input digital speech signal into word series 50 by utterances and imparts part-of-speech information 52, a sound likelihood score 54, a language likelihood score 55, a word likelihood score 56, a word continuance time length 58, the number 60 of phonemes, and a phoneme continuance time length 62 of each word to the word. An information conversion section 20 performs conversion into utterance feature quantity vectors, by utterance units 54, 55, 56, 58, 60, and 62, which include statistic values such as mean values and variance values by the utterance units and decision values by classification using the part-of-speech information 52 as elements, and a reliability imparting section 22 finds reliability based upon the degree of recognition estimated by using the vectors and an identification model previously found by learning. COPYRIGHT: (C)2007,JPO&INPIT
|
