To provide a speech recognition system which is robust to noise.
For speaker adaptation, 1st feature vector generating means (7, 8, 9) generate a feature vector series [ci,M] excluding additive noise and multiplicative noise and a 2nd feature vector generating means (12) generates a feature vector series [si,M] including additive noise and multiplicative noise. A path search part (10) makes a path search by contrasting the feature vector series [ci,M] with the standard vector [an,M] of a standard speech HMM (300) and a speaker adaptation part (11) generates an adaptive vector [xn,M] by performing correlation operation of the mean feature vector [s exp.n,M] of the standard vector [an,M] corresponding to the path search result Dv and the feature vector series [si,M] to update the feature vector of a speaker adaptive sound model (400) used for speech recognition with the adaptive vector [xn,M].
TOYAMA SOICHI
Next Patent: SPEECH RECOGNIZING METHOD