A new noise-tracking algorithm for generalizing binary time-frequency (T-F) masking to ratio masking

In this paper, we attempt to generalize the ideal binary mask (IBM) estimation to the ideal ratio mask (IRM) estimation. Under binary masking, the error in IBM estimation may greatly distort the original speech spectrum. The main purpose of this paper is using ratio mask to smooth this negative impact. Since the key issue is the noise tracking, we firstly use exponential distributions to model the distribution of noise power with binary mask and mixture power as condition. Then, we use a Gaussian distribution to model the correlation of noise estimation between adjacent T-F units. As the IBM of majority units can be estimated correctly, the correlation model could reduce the impact introduced by the error in IBM estimation. Systematic experiments show that our algorithm outperforms a common binary masking based method in terms of SNR gain and PESQ scores.