A multichannel beamforming-based framework for speech extraction

Abstract

This work proposes a new framework for multichannel speech extraction (MCSE) of one target speaker from mixtures of multiple speakers. In this framework, a beamforming technique that is based on a prior knowledge of the desired speaker's position is used. The car environment is a good example of an application where the position of the desired speaker (the driver) is known in advance. The proposed method is an optimum spatial filter with a structure inspired by the minimum variance distortionless response (MVDR) beamformer and its practical realisation through generalised side lobe canceller (GSC). The performance of the proposed method through different experiments form the 'multichannel in-car speech and noise database' must be tested and evaluated. Simulative experiments include one desired speaker (driver) and one interferer speaker (co-driver) who talked simultaneously in car environment. The experimental results show that the proposed method provides satisfactory performance.