| 摘要 |
<p>PROBLEM TO BE SOLVED: To provide a voice recognition system which is robust with respect to multiplicative distortion and additive noise generated. SOLUTION: An initially combined HMM 16 is generated from a voice HMM 10 having multiplicative distortion and an initial noise HMM of additive noise, and a Jacobean matrix [J]ij is computed by a Jacobean matrix computing section 19. Then, the Jacobean matrix is multiplied to a noise fluctuation component Namh(cep) that is obtained by combining an estimated value Ha(cep) of multiplicative distortion and obtained from actually uttered voice, an adding nature noise Na(cep) obtained in a non-uttered interval and an additive noise Nm(cep) of the noise HMM 17. The result of the multiplication and the initially combined HMM 16 are combined to generate an adaptive HMM 26. Having conducted above operations, the HMM 26 which is compatible with an observation value group RNa(cep) generated from actual uttered voice can be beforehand generated. Thus, during the voice recognition conducted by collating the group RNa(cep) and the HMM 26, adverse effects caused by the multiplicative distortion and additive distortion are canceled, an effect equivalent to the voice recognition employing clean voice is obtained and a robust voice recognition system is realized.</p> |
