Jouji Suzuki
Miyuki Seto
Tetsuya Shimamura
Dept. of Inf. and Comput. Sci., Saitama Univ., 255 Simo-okubo, Urawa, 388 Japan
Precise and correct measurement of fundamental frequency F0 is requested in the detailed analysis of speech sound and analysis synthesis telephony employing F0. This paper proposes a new method to estimate F0 employing the harmonic number and its frequency. At first, 256 pts. of speech samples are windowed and ``0''s are added to 1024 pts. The magnitude spectrum is calculated by FFT and smoothing is carried out by moving average of 5 pts. Such procedures make it easy to identify the spectrum peaks and improve the accuracy of their frequency. Every harmonic is identified from the peak of the spectrum and an harmonic number is assigned. Now fundamental frequency is calculated by (nth) harmonic frequency)/n. Finally the fundamental frequency is computed as the weighted mean, which is proportional to n, on harmonics located under 2 kHz. This idea is quite similar to getting F0 as the coefficient of a regression line passing through the origin. The accuracy of F0 obtained by this method is expected to be less than 3 Hz in female voice when the sampling frequency is 10 kHz. Experimental results indicate that the proposed method gives more precise and reasonable pitch information than the cepstrum or correlation method does.