Abstract
Conventional linear prediction (LP) analysis is known to suffer from problems in estimating the formant frequencies (vocal tract resonances) of high-pitched speech signals. The performance of conventional LP analysis deteriorates due to the harmonic structure of the glottal excitation source, especially in the case of high-pitched speech signals. Attempting to resolve this problem, a pitch-synchronous analysis technique based on a short-time energy function is presented. The proposed method has been verified to reduce the effect of the harmonic structure of the glottal excitation source. Experiments were carried out using synthetic vowels and real vowels. The results show that the proposed method yields a better performance in the estimation of formant frequencies than some previous LP analysis methods.
