自乗音声波のピーク検出による基本周波数抽出

抄録

This paper describes a method to extract fundamental frequencies of speech signals for use in speech analysis-synthesis systems. The method is based on peak-picking of squared and filtered speech waveforms. The following two filter characteristics are tried : (1) sin_2(ωt/2)/(ωt/2). (2) jωa/(jω+a)_2. Both filters are used in cascade with a low pass filter the unit response curve of which is the same as the Hamming window function. A new peak-picking method and a new period decision logic are also developed. In order to evaluate this method, gross pitch errors, unvoiced-to-voiced errors and voiced-to-unvoiced errors are calculated for the two filter characteristics. A_nA/B preference test is also conducted on speech samples which are LPC analyzed and synthesized. A female voice, a male voice and a telephone-band-limited male voice are used as the test samples. It is concluded that the method proposed here has higher preference score than the cepstral method, and that in the case of the telephone-band-limited voice, the second filter is preferable to the first one.