摘要 |
PROBLEM TO BE SOLVED: To improve recognition performance with a small study data, to expect further improvement in performance if a large amount of data are collected, and moreover, to eliminate users' load for adaptation as much as possible. SOLUTION: In a speaker adaptation mode, a phoneme label group decision part 13 requests for correct answer phoneme group information concerning an input speech of a specific speaker by collation with HMM in a dictionary storage part 15 corresponding to a correct answer phoneme group, and also requests for best phoneme group information for a score to become a maximum by collation with all HMM(Hidden Markov Model) in a dictionary storage part 15. An adaptation part 14 studies an average vector and a variance of a phoneme HMM in the dictionary storage part 15 based on the maximum posterior probability presumption method according to a correct answer phoneme group information adaptation, and further, extracts a speech pattern where a phoneme label different from the correct answer phoneme label is allocated, by comparing a phoneme label group in the correct answer phoneme group information with that in the optimal phoneme group information, and subtracts the speech pattern from the average vector of phoneme HMM corresponding to the phoneme label. |