Reproduce “User-guided Generative Source Separation” 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
Yutong Wen, Minje Kim, Paris Smaragdis
GuideSep introduces a conditional diffusion model for flexible, instrument-agnostic music source separation, leveraging user-provided mimicry (e.g., humming) and mel-spectrogram masks. The approach demonstrates superior objective and subjective separation quality compared to a predictive baseline, addressing a long-standing challenge of extracting arbitrary instruments from complex musical mixtures.
Key insights
- Generative diffusion models, specifically GuideSep, yield significantly superior perceptual quality and artifact reduction in source separation compared to a predictive baseline, even when objective SDR differences are modest.
- User-provided waveform mimicry and mel-spectrogram masks offer highly effective and flexible conditioning for instrument-agnostic source separation, enabling users to isolate specific, arbitrary instruments.
- The model's ability to generalize to real-world humming inputs (demonstrated on the HumTrans dataset) highlights the practicality and robustness of the mimicry-based guidance mechanism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.