ESTIMATION OF SPEECH COMPONENTS BY ACF ANALYSIS IN A NOISY ENVIRONMENT

A speech signal can be decomposed into the fundamental frequency and harmonics, and the autocorrelation function (ACF) is an effective tool for identifying the fundamental frequency and the harmonics. This paper, thus, explains how ACF harmonic analysis can be applied to speech detection and reconstruction when speech communication technologies are used in noisy environments. The dominant sinusoidal components used for the ACF analysis can be picked out from the short-time Fourier spectrum records of a noisy speech signal by using a peak-picking method. Because the number of components usable for speech reconstruction depends on the signal-to-noise (S/N) ratio, we authors developed new methods for peak-picking method and for harmonic sieving. The number of components picked out is adjusted frame by frame depending on the short-time S/N ratio, and harmonics are extracted from the short-time Fourier spectrum record by changing the frame length adaptively according to the fundamental frequency. Consequently, intelligible speech without “musical noise” could be reconstructed from noisy speech signals.