Improving the Quality of a Noisy Speech Signal

Abstract

In this paper we discuss the problem of reducing the noise level of a noisy speech signal. Several variants of the well-known class of “spectral subtraction” techniques are described. The basic implementation consists of a channel vocoder in which both the noise spectral level and the overall (signal + noise) spectral level are estimated in each channel, and the gain of each channel is adjusted on the basis of the relative noise level in that channel. Two improvements over previously known techniques have been studied. One is a noise level estimator based on a slowly varying, adaptive noise-level histogram. The other is a nonlinear smoother based on inter-channel continuity constraints for eliminating the so-called “musical tones” (i.e., narrow band noise bursts of varying pitch). Informal listening indicates that for modest signal to noise ratios (greater than about 8 dB) substantial noise reduction is achieved with little degradation of the speech qualify.