Adaptive speech enhancement for speech separation in diffuse noise

An adaptive enhancement method is proposed to improve recognition accuracy on the outputs of blind speech separation (BSS) system based on adaptive decorrelation filtering (ADF) in diffuse noise. A divide and conquer strategy is taken to deal with the noise effects on both system adaptation and ADF outputs. First, fast noise compensation (NC) is performed for filter adaptation, forcing ADF to focus on the task of separation; then, output noises are reduced by conventional speech enhancement, such as spectral subtraction or subspace methods. To make stationary-noise reduction techniques fit for output noises with time-varying properties caused by ADF adaptations, a fast adaptive procedure is developed to map known stationary input noise statistics to output. Separation and recognition experiments were conducted for both real and simulated diffuse noises, based on TIMIT speech data and impulse response data from a room with reverberation time T60 = 0.3sec. The proposed techniques significantly improved phone recognition accuracy of ADF results.