SPEAKER ADAPTATION DEVICE AND SPEECH RECOGNIZER

摘要

PURPOSE: To provide a speaker adaptation device capable of preventing erroneous adaptation learning even in the case of less adaptive data and adapting a speaker model so as to allow the execution of speech recognition by preventing the erroneous recognition in speech recognition over plural phonemes. CONSTITUTION: The speaker model of an HMM 11 for speech recognition is calculated by using a movement vector indicating a relation of the characteristic vectors of a hidden Markov model before and after speaker adaptation and making learning by speaker adaptation of the initial speaker model in accordance with the learning data for speaker adaptation. The first characteristic vector of the HMM 11 after speaker adaptation based on the learning data for speaker adaptation when the learning data exists is subjected to smoothing processing using the plural second characteristic vectors of the HMM 11 after the speaker adaptation existing near the same. The second characteristic vectors which do not exist near the first characteristic vector are exempted in accordance with the neighboring state between the first characteristic vector and the plural second characteristic vectors.