Spectral Subtraction With Variance Reduced Noise Spectrum Estimates

Abstract

Spectral subtraction has the drawback that it introduces an unpleasant residual noise. This noise is a result of under-subtraction which occurs due to high variance of noise magnitude spectrum estimates. In this study we investigate a number of smoothing techniques that can be employed to reduce this variability. We extend the scope of this paper by using the phase spectrum in a novel manner along with the processed magnitude spectrum. This is based on recent findings which suggest that estimation of the phase spectrum using low dynamic range analysis windows (at 20-40ms window durations) is beneficial for speech enhancement. Using an objective speech quality measure and spectrogram analysis we show that the smoothing of noise magnitude spectrum estimates is an effective method of suppressing musical noise. We also show that the use of a low dynamic range analysis window for estimation of the phase spectrum of noisy speech results in a reduction of background noise. However, we found that combining the two techniques offers no advantage over spectral subtraction alone.