| 摘要 |
PURPOSE:To improve the performance of recognition, to shorten a calculation time and to reduce an adaptive learning voice. CONSTITUTION:The hidden Markov model(HMM) of continuous mixture probability distribution for unspecified speakers is formed from many speaker's voices in each category to be recognized, the learning voice of a speaker A is inputted and only the divided weight feature of an unspecified speaker HMM is optimized so that the tolerance of the voice is maximized. When a certain mixed distribution HMM for an unspecified speaker is the synthesized distribution 4 of acoustic feature distribution 1 to 3, the adapting result of the speaker A to the voice is the average value of respective mean values of the distribution 1 to 3, its covariance is the same, but only it weight coefficient is changed by adaptation. Namely the distribution 2 is especially increased and distribution 1, 3 are reduced. |
