Binary mask estimation for noise reduction based on instantaneous SNR estimation using Bayes risk minimisation

The binary mask approach has been researched to suppress noise and improve speech intelligibility in noisy environments. An algorithm that estimates the binary mask for noise-corrupted speech based on the instantaneous signal-to-noise ratio (SNR) estimation is proposed. The instantaneous SNR estimation is performed by minimising the Bayes risk with a weighted cost function. In the experiments, white noise was used for the training of the SNR estimator and the binary mask estimation was performed for babble, factory, speech-shaped noise. The experimental results show that the proposed method yields substantial improvements in terms of classification accuracy for the binary mask estimation.