| 摘要 |
PROBLEM TO BE SOLVED: To apply all HMM parameters with high accuracy from a single prediction even when there are not many words, by providing an adding part for calculating a parameter of a specific speaker's hidden Markov model(HMM) and an SD-HMM storing part for storing the specific speaker's HMM, etc., for this speech recognizing device. SOLUTION: A prediction parameter storage part 7 stores beforehand a prediction parameter determined by pre-learning. A predicting part 8 calculates a linear sum of a 1st difference and the prediction parameter, and a 2nd difference storing part 9 stores the 2nd difference outputted from the predicting part 8. The adding part 10 adds the 2nd difference to a parameter of SI-HMM, and outputs the sum as a speaker-adapted specific speaker's HMM of a new speaker. An SD-HMM storing part 11 stores the specific speaker's HMM outputted from the adding part 10. This specific speaker's HMM stored in the SD-HMM storing part 11 is the HMM for the new speaker obtained as a result of speaker adaptation by a small amount of utterance of the new speaker for adaptation. |
