Encoder Re-training with Mixture Signals on FastMVAE Method

In this paper, we propose a new network training to improve the source separation performance of the fast multichannel variational autoencoder (FastMVAE) method. The FastMVAE method is very effective for supervised source separation. It also significantly reduces the processing time by replacing the backpropagation steps in the MVAE method with a single forward propagation of the encoder for estimating latent variables. In previous studies, the encoder is trained together with the decoder using clean speech. In contrast, in this study, we re-train only the encoder by using the mixed signals with the decoder fixed. More specifically, using the imperfectly separated signals obtained in the process of the source separation algorithm, we train the encoder to find the optimal latent variables that minimize the objective function for source separation. Experimental results show that the proposed method reduces the objective function at almost every iteration and achieves higher separation performance than the conventional method.