Speech Enhancement and Noise Reduction

10.10 Summary

Speech signals, when propagating through acoustic channels and recorded by microphone receivers, are inevitably corrupted by noise and interference. How to estimate the speech of interest from its corrupted observations has become one of the most challenging problems in acoustic signal processing, which involves a wide variety of techniques such as source separation, channel identification, speech dereverberation, to name a few. This chapter was devoted to the speech-enhancement and noise-reduction problem, which primarily aims to recover the desired speech signal from its realizations corrupted by additive noise. It covered not only the well-recognized single-channel techniques such as Wiener filtering and spectral subtraction, but also the techniques using two or more microphone sensors. In the former case, the study demonstrated that noise reduction is possible, but at a price of distorting the clean speech, and more noise reduction corresponds to more speech degradation. Although some schemes can be arranged to restrain the speech distortion so that it will not be perceived by listeners, they in general lead to less noise reduction. This performance trade-off problem, however, becomes less serious and can even be avoided when two or more microphone sensors can be used. In this situation, both the adaptive-noise-cancellation and beamforming techniques can be employed to mitigate the unwanted noise effects with less or even no speech distortion.