A modulation model for the vowel sound /i/ in human speech

Abstract

An approach has been developed that models a nonstationary speech signal with very low error on a pitch period by pitch period basis using an AM–FM model. The speech signal is first separated into two independent signals, a low-frequency signal and a high-frequency signal. AM and FM signals are extracted from the high-frequency signal through the use of two nonlinear operators, the integral operator and the overmodulation detection algorithm. Model parameters are determined by a combination of direct Fourier series computation and parameter optimization using the correlation error measure and the Marquardt–Levenberg algorithm. This technique was applied to speech produced in various gas concentrations of air and helium in order to determine whether model parameters are directly related to the velocity of sound ratio. It was found that the amplitudes of the harmonics of the low-frequency signal were emphasized with the increase of the velocity of sound ratio. The standard deviation of the carrier frequency of the FM signal tended to increase with the velocity of sound ratio. The amplitudes of the AM signal had a modal structure, and the modes were independent of velocity of sound ratio.