Unsupervised Domain Adaptation of Universal Source Separation Based on Neural Full-Rank Spatial Covariance Analysis

This paper presents an unsupervised domain adaptation method of universal source separation (USS) using a neural blind source separation (BSS) technique. USS is a task of separating environmental sound mixtures and has been investigated by training a deep neural network in a supervised manner. This approach, however, suffers from the domain mismatch problem between training and test data. In this paper, we propose an unsupervised domain adaptation method based on neural full-rank spatial covariance analysis. In the (pre-)training phase, a neural separation model is trained to perform multichannel Wiener filtering in a supervised manner. In the test phase, on the other hand, the network is fine-tuned to fit the test mixtures in an unsupervised manner by maximizing a log-marginal likelihood of a nonlinear (neural) BSS model. Experimental results with simulated mixture signals show that our method successfully improves the separation performance for test sets without any additional supervision.