| 摘要 |
PURPOSE: To increase the discrimination accuracy of phonemes by inputting an input speech signal to a band pass filter(BPF) group consisting of plural channels and inputting the signal also to a voiced/voiceless sound judging means and a pitch extraction part. CONSTITUTION: An input speech signal is inputted to the BPF group, and when a zero crossing interval(ZCI) measured from respective channels and an output waveform from the BPF group is included within a prescribed range, the value of a variable expressing the generation frequency of the ZCI in the prescribed range is increased by a prescribed value to find out the distribution of variables. When the input voice signal is judged as a voiced sound by the judging means, the feature value of the input voice signal is found out based upon a channel having center frequency corresponding to the integer times of reference frequency outputted from the pitch extraction part and a variable value in a range corresponding to the integer times of the reference frequency out of the variable value distribution obtained from the BPF group. In the case of a voiceless sound, the feature value of the input voice signal is found out based upon a variable value in a range corresponding to each center frequency out of the variable value distribution from the VPF group. |
