Blind Source Extraction Based on Multi-channel Variational Autoencoder and X-vector-based Speaker Selection Trained with Data Augmentation

Multi-channel variational autoencoder (MVAE) has been proven to be a promising method for blind source separation (BSS), which is an elegant combination of the strong modeling power of deep neural networks (DNN) and interpretable BSS algorithm based on independence assumption. However, the success of MVAE is limited to the training with very few speakers and the speeches of desired speakers are usually included. In this paper, we develop a sequential approach for blind source extraction (BSE) by combining MVAE with the x-vector based speaker recognition (SR) module. Clean speeches of 500 speakers are utilized in training MVAE to verify its capability of generalization to unseen speakers, and an augmented dataset is constructed to train the SR module. The efficacy of the proposed BSE approach in terms of extraction accuracy, signal-to-interference ratio (SIR) and signal-to-distortion ratio (SDR) are validated with test data consisting of unseen speakers under varied environments.