Low residual noise speech enhancement utilizing time-frequency filtering

Spectral subtraction is a well known technique for enhancing speech corrupted by additive wideband noise. In this technique, the “clean” signal is approximated by subtracting a noise estimate from the spectrum of the corrupted signal. A negative side effect is the residual “musical” noise that is produced when isolated spectral peaks exceed the noise estimate. In this paper, a low residual noise enhancement method is presented. This method is based on spectral subtraction but incorporates an algorithm developed to suppress “musical” noise without affecting speech. The algorithm is referred to as time-frequency filtering because spectral peaks due to noise are eliminated on the basis of duration, bandwidth, and proximity to other peaks. Results showing the effects of combining spectral subtraction and time-frequency filtering are given