TSEGAN:Target Speech Extraction Algorithm Based on Generative Adversarial Networks

Single-channel speech extraction algorithms based on deep learning need to calculate time-frequency mask, but the time-frequency mask function in existing speech extraction methods needs to be manually selected, and it is not encapsulated into a deep network for optimization learning, usually relying on Wiener filtering method for subsequent processing. In this paper, we proposes a speech extraction algorithm based on generative adversarial network. In the generator part, a learnable time-frequency mask generation algorithm is used. A recursive derivation algorithm and a sparse encoder are introduced to improve the generation result of time-frequency mask. We use the mixed audio to replace the random noise in the original GAN model as input, to increase the correspondence of the network model topology in the form of latent variables, and improve the fitting ability of the generative model. Compared with the basic algorithm, the SDR and SIR separation indexes of the proposed method increase by 5dB and 5.5 dB.