Enhancement of highly noisy speech signals

Abstract

The speech enchancement problem is divided into two basic steps: identification of speech and non-speech regions and noise cancellation. Non-conventional fuzzy set operations are used to identify the speech and noise regions in a noisy speech signal. A careful study of the sub-band energies is done to get a sound idea about the nature of noise and its variation. A neural net is trained adaptively with the noisy speech and the neighbouring noise regions. A Sugeno type fuzzy inference system is also used for the purpose of noise cancellation. Finally the signal is band-pass filtered and played back by relatively varying the amplitudes in the speech and non-speech regions, producing a more intelligible speech.