SPEAKER NORMALIZING PROCESSOR AND VOICE RECOGNITION DEVICE

摘要

PROBLEM TO BE SOLVED: To generate an acoustic model and to recognize a voice by highly accurately estimating a frequency warping function, executing speaker normalization and learning an initial HMM. SOLUTION: A speaker normalization processing part 12 estimates the featured value of a vocal path shape to be the anatomical shape of each learning speaker's vocal path based on the vocal waveform data of each learning speaker while referring to correspondence relation between a vocal path shape parameter determined based on the vocal path model of a reference speaker and formant frequency, and also estimates the formant frequency based on the estimated result and generates a frequency warping function by straight interpolating the estimated formant frequency and the formant frequency corresponding to the reference speaker in relation to the frequency warping function. After applying speaker normalization to the voice waveform data of each learning speaker by using the frequency warping function, the processing part 12 extracts an acoustic feature parameter and learns an initial HMM based on text data corresponding to the acoustic feature parameter to generate a normalized HMM. Thus voice recognition is executed by using the HMM.