モーメント計算によるホルマント周波数の抽出

抄録

This paper describes the method of extraction of formant frequency from digitalized speech spectra by a computer program and its application to natural speech materials. Moment calculation method was employed for the formant frequency extraction, paying special attention to the separation of the first from the second, and the second from the higher formant frequency ranges. The parameters F^^^- and F_&lt2e&gt were used in this separation, where F^^^- is the first order moment of the speech spectrum and F_&lt2e&gt is the ratio of the second order moment to the first order moment of spectrum. After properly separating frequency ranges of each formant, the formant frequencies are calculated as the first order moment (average frequency) within these separated portion of the spectrum. The comparison of formant frequencies extracted from natural speech with corresponding spectrograms and the results of analysis of errors proved that this method gives acceptable results in actual speech analysis. The results can be regarded as the first approximation of formant frequencies and will form the basis of more detailed analysis such as "analysis by synthesis" or of future study on segmentation and recognition of Japanese monosyllables.