Reproduce “GASS: Generalizing Audio Source Separation with Large-scale Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jordi Pons, Xiaoyu Liu, Santiago Pascual, Joan Serrà
Researchers at Dolby Laboratories developed a single deep learning model for general audio source separation (GASS), capable of simultaneously separating speech, music, and sound events. This was achieved by training on an unprecedented 15,000+ hour dataset, demonstrating strong generalization and transfer learning capabilities across diverse audio tasks.
Key insights
- A single, unified deep learning model can effectively perform general audio source separation across diverse audio types: speech, music, and sound events.
- Leveraging large-scale supervised training (over 15,000 hours) is crucial for achieving robust generalization in complex audio source separation tasks.
- Pre-trained General Audio Source Separation (GASS) models serve as strong foundational models, enabling robust transfer learning even to tasks with different output definitions and consistently outperforming models trained from scratch.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.