Dr. Patrick A. Naylor

Dereverberation of speech signals is an important speech enhancement process for hands-free terminals. For any non-close-talking microphone, the observed signal represents the original speech convolved with an unknown, time-varying acoustic channel, together with additive noise. The convolution can cause the speech to sound distant, spectrally distorted and can reduce intelligibility. Dereverberation aims to undo the effects of the acoustic channel between the talker and the microphone. This is a blind problem and currently an unsolved problem.

This presentation aims to define the problem, review existing approaches and discuss current work on this topic. Broadly speaking, existing approaches fall into two categories - one based on processing of the LPC prediction residual and one based on a combination of blind channel estimation and channel inversion. Examples will be presented in order to compare approaches in these two categories. The quantitative and qualitative measurement of reverberation will also be discussed. Some of the results of new work on multi-channel adaptive algorithms for blind channel estimation will be presented and the strengths and weaknesses of such techniques will be set in the context of the overall aim of dereverberation.

Full Paper