Reproduce “Less is More: Data Curation Matters in Scaling Speech Enhancement” on OpenResearch

About this paper

Chenda Li, Wangyou Zhang, Wei Wang, Robin Scheibler, Kohei Saijo, Samuele Cornell, Yihui Fu, Marvin Sach and 5 more

Researchers, including those from Shanghai Jiao Tong University and Carnegie Mellon University, demonstrated that carefully curated, smaller datasets for speech enhancement consistently outperform larger, uncurated ones, particularly in improving perceptual quality. The study, motivated by observations from the URGENT 2025 challenge, found that a 700-hour quality-filtered subset yielded better non-intrusive metric scores than a 2,500-hour uncurated dataset, and a 100-hour top-ranked subset even surpassed the full 2,500-hour set in some perceptual evaluations.

Key insights

• Imperfections in supposedly 'clean' speech labels are a critical, often overlooked factor that limits performance gains from increasing data quantity in speech enhancement.
• Data quality, as measured by non-intrusive metrics, is more crucial than sheer data volume for achieving robust perceptual quality improvements in speech enhancement systems.
• A combined strategy of warm-starting models on broad, large datasets and then fine-tuning on meticulously curated, high-quality data yields superior overall performance.

Summary by alphaXiv.