Towards Reliable Objective Evaluation Metrics for Generative Singing Voice Separation Models

This upload accompanies the WASPAA 2025 paper, 'Towards Reliable Objective Evaluation Metrics for Generative Singing Voice Separation Models'.

It contains the evaluation audio used to compute the evaluation metrics, as well as the loudness-normalised test stimuli used in the DCR test.

Also included is a CSV file containing all the metrics and DMOS scores used to calculate Spearman's rank-based correlation coefficients (SRCCs) for evaluating both discriminative and generative models.
A demonstration Python script outlines how the SRCCs between the DMOS and the evaluation metrics for both types of model are calculated.

The Readme.md file provides step-by-step instructions on how to benchmark metrics not included in the analysis.