A new spectrum estimation method of a speech waveform by eliminating high frequency pitch period

This paper proposes a new spectrum estimation method for speech signals which are uttered with a short pitch period. This method is one of the excellent spectrum estimators in which an exact spectrum of a linear model can be obtained even if quite high frequency pitch pulses are fed into the model as an input. This method consists of two-stage analyses. The troughs of harmonics are removed by using cepstrum analysis repeatedly in the first stage. After that, the spectral envelope is estimated by using the autocorrelation method in the second stage. This report shows that the proposed method can estimate the formant frequencies of speech signals uttered by children quite accurately, more so than the conventional linear prediction method