Automatic adaptive speech separation using beamformer-output-ratio for voice activity classification

摘要: This paper focuses on the practical challenge of adaptation control for speech separation systems. Adaptive beamforming methods, such as minimum variance distortionless response (MDVR), can effectively extract desired signal from interference and noise. However, to avoid cancellation problem, beamformer is halted when speaker active. An automated scheme this requires classifying speakers' voice activity status, which remains a multi-speaker environments. In paper, we propose novel approach identify activities two speakers based new metric, called beamformer-output-ratio (BOR). Statistical properties BOR are studied used develop hypothesis-based method classification. The further refined using an algorithm detecting incorrect by analysing changes in output power blind adapting MVDR beamformer. Based construct automatic adaptive system simultaneously separate speakers. module uses beamformers whose guided Our methods lead to, some cases, 20% reduction classification error, 8dB improvement SINR. results verified both synthesised signals realistic recordings. HighlightsWe design speakers.The quantity its roles active identification introduced.The BOR-VAC developed, generic form realisation.We model behaviour detect adaptation.The proposed systems tested real