Distant speech recognition in reverberant noisy conditions employing a microphone array

This paper addresses the problem of distant speech recognition in reverberant noisy conditions employing a microphone array. We present a prototype system that can segment the utterances in real-time and generate robust ASR results off-line. The segmentation is carried out by a voice activity detector based on deep belief networks, the speaker localization by a position-pitch plane, and the enhancement by a novel combination of convex optimized beamforming and vector Taylor series compensation. All of the components are compared with other similar ones and justified in terms of word accuracy on a proposed database which simulates distant speech recognition in a home environment.