Speech Denoising in Multi-Noise Source Environments Using Multiple Microphone Devices Via Relative Transfer Matrix

Speech denoising is a challenging problem when there are multiple active noise sources. This paper introduces a novel blind denoising approach using the Relative Transfer Matrix (ReTM) as a spatial feature of noise source locations and the environment in multi-microphone settings. The ReTM is a generalization of Relative Transfer Function (ReTF) for simultaneously active sources and multiple receivers. We allocate receivers into two multichannel groups and formulate the ReTM to describe the spatial mapping between them. The ReTM with respect to noise sources is estimated blindly using covariance matrices of microphone recordings during speech-free intervals. We use the ReTM to estimate the noise at one group of microphones from the other. The estimated noise is then subtracted from the incoming signal to achieve speech denoising. We illustrate the effectiveness of the proposed algorithm through simulations and experimental recordings. The method does not require prior knowledge of the number of speech and noise sources, nor source and microphone locations, and can be extended to a configuration with more than three microphones.