On loss functions and evaluation metrics for music source separation

TL;DR

This paper systematically benchmarks various loss functions for music source separation, evaluates their effectiveness as metrics, and highlights the limitations of standard evaluation methods, proposing alternatives based on these losses.

Contribution

It provides a comprehensive comparison of loss functions for music source separation and explores their potential as evaluation metrics, addressing limitations of current standards.

Findings

Certain loss functions outperform traditional metrics in separation quality.

Some loss-based metrics correlate better with subjective listening tests.

Standard SNR metrics can be misleading in specific scenarios.

Abstract

We investigate which loss functions provide better separations via benchmarking an extensive set of those for music source separation. To that end, we first survey the most representative audio source separation losses we identified, to later consistently benchmark them in a controlled experimental setup. We also explore using such losses as evaluation metrics, via cross-correlating them with the results of a subjective test. Based on the observation that the standard signal-to-distortion ratio metric can be misleading in some scenarios, we study alternative evaluation metrics based on the considered losses.