A New Weighted Loss for Single Channel Speech Enhancement under Low Signal-to-Noise Ratio Environment

This work studies the single channel speech enhancement problem in the case of low signal-to-noise ratio (SNR). To that aim, the supervised learning technique is utilized, where a new loss is developed to trade-off the speech distortion and residual noise. By a use of weighted combination of distortion and residual noise, the noise suppression and speech quality are considered simultaneously. In doing so, it also is easy to verify that the commonly used mean square error (MSE) loss is a special case of the proposed loss. Experimental results show, with the convolutional encoder-decoder-long short-term memory (CED-LSTM) network, the proposed loss outperforms the MSE and the recently proposed scale-invariant signal-to-distortion ratio (SI-SDR) loss.