A reverberation robust target speech detection method using dual-microphone in distant-talking scene

Abstract

Speech signal processing with coherent interference in reverberant environment under distant-talking scene has always been a difficult problem, in which Target Speech Detection (TSD) plays a basic role. This paper proposes a reverberation robust TSD method, which is based on Beam-to-Reference Ratio (BRR) using a dual-microphone array, is proposed. At first, detection thresholds in Time–Frequency (T–F) domain are derived under free sound field assumption. A novel estimator, Direct-to-Reverberate Ratio (DRR), is introduced to enlarge the basic assumption to reverberant environment which is common in distant-talking scene. Then T–F thresholds of BRR are revised according to DRR. Meanwhile, innate weak point of compact array is studied due to spatial aliasing and a sidelobe suppression procedure is proposed to further eliminate the effect of the coherent interference. According to these techniques, a state-of-the-art full-band judgement is obtained by statistics of likelihood on each T–F block. Experimental results show that the proposed method performs robust in different reverberant environments with coherent interferences when target speech is from the a priori known direction-of-arrivals (DOA) in distant-talking scene.