Abstract: Improving the quality of separated speech under low SNR is the focus of speech separation technology research, while most methods still only train the features of the target speaker's speech under low SNR. Aiming at the shortcoming of current methods, a mixed speech separation method based on cooperative training generative adversarial networks(GAN) is proposed. In order to avoid the extraction of the complex acoustic feature, the generative model uses a fully convolutional neural network to directly extract the high-dimensional features of the time-domain waveform, and the discriminative model obtains the features of the interference speaker by constructing a binary classification convolution neural network. Then, the source of the separated information obtained by the system is no longer single. Experiments show that the proposed method can better recover the information of high-frequency components under low SNR, and the separation performance is better than that of the comparative methods on the two-speaker mixed speech data set.

Key words: low SNR, generative adversarial networks, cooperative training, speech separation ,