摘要 |
PROBLEM TO BE SOLVED: To improve speech recognition performance by adapting error correction model to an utterance content.SOLUTION: A speech recognition section 11 recognizes a speech included in a voice data using an error correction model currently stored in an error correction model storage 23. An error correction section 12 corrects a correct text candidate obtained as a result of the speech recognition according to input made by a user to generate a correct word string. An alignment section 13 aligns the words included in the correct word string in time sequence based on the voice data. A feature amount extraction section 14 extracts linguistic characteristics from the correct text candidate and the aligned correct word string. A model parameter learning section 15 calculates a feature weight for a feature function for processing linguistic characteristics as a rule and a mixture weight of plural language models constituting a mixture model in a statistical manner based on the extracted linguistic characteristics, the correct text candidate, and an acoustic score of the aligned correct word string and a language score to update the error correction model. |