Improving Source Extraction with Diffusion and Consistency Models

Tornike Karchkhadze, Mohammad Rasool Izadi, Shuo Zhang

Keywords: source extraction, consistency models, score-matching diffusion

TL;DR: We integrate a score-matching diffusion model with consistency distillation into a deterministic architecture for time-domain musical source extraction, significantly improving performance on the Slakh2100 dataset.

Abstract: In this work, we integrate a score-matching diffusion model into a standard deterministic architecture for time-domain musical source extraction. To address the typically slow iterative sampling process of diffusion models, we apply consistency distillation and reduce the sampling process to a single step, achieving performance comparable to that of diffusion models, and with two or more steps, even surpassing them. Trained on the Slakh2100 dataset for four instruments (bass, drums, guitar, and piano), our model shows significant improvements across objective metrics compared to baseline methods. Sound examples are available at https://consistency-separation.github.io/.

Submission Number: 25