Speech enhancement based on a sinusoidal model.

Abstract

Sinusoidal modeling is a new procedure for representing the speech signal. In this approach, the signal is divided into overlapping segments, the Fourier transform computed for each segment, and a set of desired spectral peaks is identified. The speech is then resynthesized using sinusoids that have the frequency, amplitude, and phase of the selected peaks, with the remaining spectral information being discarded. Using a limited number of sinusoids to reproduce speech in a background of multi-talker speech babble results in a speech signal that has an improved signal-to-noise ratio and enhanced spectral contrast. The more intense spectral components, assumed to be primarily the desired speech, are reproduced, whereas the less intense components, assumed to be primarily background noise, are not. To test the effectiveness of this processing approach as a noise suppression technique, both consonant recognition and perceived speech intelligibility were determined in quiet and in noise for a group of subjects with normal hearing as the number of sinusoids used to represent isolated speech tokens was varied. The results show that reducing the number of sinusoids used to represent the speech causes reduced consonant recognition and perceived intelligibility both in quiet and in noise, and suggests that similar results would be expected for listeners with hearing impairments.