Study examines SI-SDR metric in speech separation with noisy references

Arxiv CS.SD (Sound/Audio ML)

Paper investigates implications of using Scale-Invariant Signal-to-Distortion Ratio (SI-SDR) as evaluation and training objective in supervised speech separation when training references contain noise. Noisy references impact the metric's effectiveness.

voice-ai

Arxiv CS.SD (Sound/Audio ML)·Simon Dahl JepsenMads Gr{\ae}sb{\o}ll ChristensenJesper Rindom Jensen

7 days ago