Speech enhancement using a soft-decision maximum likelihood noise suppression filter

One way of enhancing speech in acoustic noise environments is to process the data using a noise suppression filter which performs a spectral decomposition of a frame of noisy speech and attenuates a particular spectral line depending upon how much the measured speech plus noise power exceeds an estimate of the background noise. In this paper a two-state model (speech absent or speech present) is used in the derivation of a maximum likelihood estimator for the speech power. The result is a suppression curve that is similar to existing algorithms but which is weighted by the a posteriori probability that the current measurement corresponds to the speech state. A class of curves is obtained by varying the value of a suppression factor which can be chosen to trade off noise suppression against speech distortion. The algorithm has been implemented in real time in the time domain, exploiting the structure of the channel vocoder to perform the spectral decomposition. Extensive testing has shown that the noise can be made imperceptible by proper choice of the suppression factor. A tape recording will be played demonstrating the noise suppression properties of the filter and demonstrating how the algorithm can be used in tandem with a narrow-band vocoder, such as LPC, to enhance noisy speech. [Work sponsored by Department of the Air Force.]