Single-channel Speech Separation Based on Separated SI-SNR Regression Estimation and Adaptive Frequency Modulation Network

In practical applications, speech separation models are often disturbed by unknown noise, resulting in serious degradation of generalization performance. To solve this problem, Single channel speech separation method based on separate SNR regression estimation and adaptive frequency modulation network is proposed. Firstly, the scale invariant SNR of test signal separation results is estimated by prediction network to calculate the cognitive uncertainty of the model; Then, an adaptive frequency modulation network is designed to adjust the spectrum of signals with high uncertainty to reduce the cognitive uncertainty of the model, so as to improve the generalization ability of the model in the face of unknown noise. The experimental results show that compared with the Conv-Tasnet, the proposed method improves the SI-SNR (Scale Invariant SNR) from 2.72 dB to 4.57 dB, with an increase of 67.94%, and has a great improvement in generalization ability. Compared with Conv-Tasnet with Soft-Mask, the SI-SNR is increased from 3.32 dB to 4.57 dB, with an increase of 37.65%, indicating that this method has better generalization ability than soft mask mechanism. It effectively alleviates the serious degradation of generalization ability of speech separation network in the face of unknown noise.