APPARATUS AND METHOD FOR EXTRACTING SPEAKER'S FEATURE, SPEECH RECOGNITION DEVICE, SPEECH SYNTHESIZER, AND PROGRAM RECORDING MEDIUM

摘要

PROBLEM TO BE SOLVED: To stably exact a speaker's feature from few utterance data. SOLUTION: A frequency function estimation part 5 is provided with a phonemic limit estimation part and a maximum likelihood estimation part. During learning, the phonemic sequence (language model) of contents of utterance or the language model of a feeble grammar is applied to estimate phonemic limit information by a Vitervi algorithm. Furthermore, a frequency warping function f() is estimated about a phonemic section selected on the basis of the phonemic limit information. Moreover, during recognition, the phonemic limit information is estimated by applying the language model of a feeble grammar. A frequency warpage part 4 performs the frequency warp of input sound parameter series about the phonemic section selected on the basis of the phonemic limit information. Thus, the deformation of a phoneme or an unvoiced sound part which is hardly influenced with respect to the difference of a vocal tract length is prevented by precluding the phoneme or the unvoiced sound part which is hardly influenced with respect to the difference of the vocal tract length from the subject of learning or normalization. Thereby, the speaker's feature is stably extracted from few utterance data.