| 摘要 |
<P>PROBLEM TO BE SOLVED: To provide a voice synthesizer using an HMM (Hidden Markov Model) which can suppress occurrence of distortion in a synthetic voice waveform, and an HMM learning device therefor. <P>SOLUTION: A learning device 110 comprises: a voice database 60; an FO extraction processing unit 62 for extracting a fundamental frequency (F0) from each frame of voice; an MFCC (Mel Frequency Cepstrum Coefficient) calculation unit 64 for calculating an MFCC from each frame; an MFCC conversion unit 120 for converting the MFCC into a predetermined angular amount for each frame by performing sampling of a frequency domain which makes duality with sampling of a time domain for calculating an MFCC; and an HMM learning unit 124 for performing HMM learning using the F0 and the MFCC calculated for each frame as learning data 122 and decision tree learning for selecting one of the HMMs. <P>COPYRIGHT: (C)2013,JPO&INPIT |
