SSL-Embedding-based Feature Representation for Music Source Separation

Abstract

In this paper, we propose a SSL-Embedding-based feature representation(SFR) for Music Source Separation(MSS). The proposed method generates a feature represen- tation by splitting an acoustic feature and an embedding obtained by a pretrained self-supervised learning model and aggregating them. First, SFR expands the embed- ding, and splits the acoustic feature and the extended embedding into split features. And then, SFR aggregates the split features and extracts dependencies between the aggregated features. Finally, SFR unifies the aggregated split features and obtains the feature representation that is concatenated to the acoustic feature. As a result, our proposed method was applied to an existing MSS model and showed boosted performance for separating ”bass”, ”other” and ”vocals” sources in MUSDB18 testset. In addition, It also generally improved performance in separating sources in K-CONTENTS 22 testset, which is out-of-domain dataset.