MDX-Mixer: Music Demixing by Leveraging Source Signals Separated by Existing Demixing Models

Abstract

This paper presents MDX-Mixer, which improves music demixing (MDX) performance by leveraging source signals separated by multiple existing MDX models. Deep-learning-based MDX models have improved their separation performances year by year for four kinds of sound sources: “vocals,” “drums,” “bass,” and “other”. Our research question is whether mixing (i.e., weighted sum) the signals separated by state-of-the-art MDX models can obtain either the best of everything or higher separation performance. Previously, in singing voice separation and MDX, there have been studies in which separated signals of the same sound source are mixed with each other using time-invariant or time-varying positive mixing weights. In contrast to those, this study is novel in that it allows for negative weights as well and performs time-varying mixing using all of the separated source signals and the music acoustic signal before separation. The time-varying weights are estimated by modeling the music acoustic signals and their separated signals by dividing them into short segments. In this paper we propose two new systems: one that estimates time-invariant weights using 1×1 convolution, and one that estimates time-varying weights by applying the MLP-Mixer layer proposed in the computer vision field to each segment. The latter model is called MDX-Mixer. Their performances were evaluated based on the source-to-distortion ratio (SDR) using the well-known MUSDB18-HQ dataset. The results show that the MDX-Mixer achieved higher SDR than the separated signals given by three state-of-the-art MDX models.