Title: Investigating the Scale-Invariant Signal-to-Distortion Ratio in Speech Separation Tasks Involving Noisy References

Abstract: This study explores the consequences of employing the Scale-Invariant Signal-to-Distortion Ratio (SI-SDR) as both a training objective and an evaluation metric for supervised speech separation, specifically when the reference signals used during training are corrupted by noise—a common scenario in the widely adopted WSJ0-2Mix benchmark. Mathematical analysis of the SI-SDR in the presence of noisy references indicates that such noise imposes a ceiling on the maximum achievable SI-SDR score or results in the unwanted inclusion of noise within the separated audio outputs. To mitigate these issues, the authors propose a strategy that improves the quality of reference signals and incorporates mixture data from the WHAM! dataset, with the goal of preventing models from memorizing noisy references. The performance of two models trained on these refined datasets was assessed using the non-intrusive NISQA.v2 metric. While the results demonstrate a decrease in noise levels within the separated speech, they also reveal that modifying the reference signals can introduce artifacts, thereby capping improvements in overall audio quality. Furthermore, a negative correlation was observed between SI-SDR values and perceived noisiness across various models tested on both the WSJ0-2Mix and Libri2Mix datasets, reinforcing the findings derived from the initial analysis.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC