Noise-robust Speech Separation with Fast Generative Correction

• DOI: 10.48550/arxiv.2406.07461
• Publication Date: 2024-06-11
• Authors: Helin Wang; Jesús Villalba; Laureano Moro-Vel...; Jiarui Hai; Thomas Thebaud; Najim Dehak
• Keywords: Separation (statistics); Source Separation

Speech separation, the task of isolating multiple speech sources from a mixed audio signal, remains challenging in noisy environments. In this paper, we propose generative correction method to enhance output discriminative separator. By leveraging corrector based on diffusion model, refine separation process for single-channel mixture by removing noises and perceptually unnatural distortions. Furthermore, optimize model using predictive loss streamline model's reverse into single step rectify any associated errors process. Our achieves state-of-the-art performance in-domain Libri2Mix dataset, out-of-domain WSJ with variety noises, improving SI-SNR 22-35% relative SepFormer, demonstrating robustness strong generalization capabilities.