Latent source-specific generative factor learning for monaural speech separation using weighted-factor autoencoder

Abstract

Much recent progress in monaural speech separation (MSS) has been achieved through a series of deep learning architectures based on autoencoders, which use an encoder to condense the input signal into compressed features and then feed these features into a decoder to construct a specific audio source of interest. However, these approaches can neither learn generative factors of the original input for MSS nor construct each audio source in mixed speech. In this study, we propose a novel weighted-factor autoencoder (WFAE) model for MSS, which introduces a regularization loss in the objective function to isolate one source without containing other sources. By incorporating a latent attention mechanism and a supervised source constructor in the separation layer, WFAE can learn source-specific generative factors and a set of discriminative features for each source, leading to MSS performance improvement. Experiments on benchmark datasets show that our approach outperforms the existing methods. In terms of three important metrics, WFAE has great success on a relatively challenging MSS case, i.e., speaker-independent MSS.