Blind Beamforming Using Parallel Single-channel SPeech Enhancers

Abstract

This paper presents an idea to extend a certain class of single channel speech enhancement algorithms to include the spatial domain. The resulting blind beamformer does not rely on a-priori knowledge of source and sensor positions and it enhances one or several speech sources based only on received data. The underlying principle in this approach is the fact that speech signals are short time stationary. Provided that the single channel speech enhancers attenuates unwanted sources and at the same time preserve the short time stationarity of speech signals, a summation of a small array of such single channel processors constitutes a coherent spatial speech enhancement. As opposed to traditional beamforming where the phase alteration is pre-specified, the phase alteration of the proposed structure is controlled by the received data. The evaluation uses a two microphone array and indicates that the Signal to Interference Ratio is increased for a variety of source positions using the proposed method with only an insignificant decrease in speech quality.