Research map / Label-efficient and robust learning / Domain adaptation and generalization
Unsupervised Sound Separation Using Mixture Invariant Training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add matched-domain unsupervised mixtures to mismatched supervised training | MixIT speech separation · WSJ0-2mix and Libri2Mix | SI-SNRi: 2-3 dB; about 6 dB on WSJ0-2mix and 2.5 dB for Libri2Mix | Helped |
| Train fully unsupervised using 1-or-2-source mixtures rather than only 2-source mixtures | MixIT speech separation · Anechoic and reverberant speech separation | SI-SNRi: up to 3 dB | Mixed |
| Add 10% supervised examples | MixIT speech separation · Anechoic speech separation | SI-SNRi | Helped |
| Add YFCC100m unsupervised mixtures to FUSS supervised training | FUSS supervised model with p0=0.2 · YFCC100m test set | MoMi: 0.3 -> 6.1 | Helped |
| Train fully unsupervised on YFCC100m mixtures instead of supervised FUSS | MixIT · FUSS test set | MoMi: 7.7 -> 9.6 | Helped |
| Train fully unsupervised on YFCC100m mixtures instead of supervised FUSS | MixIT · FUSS test set | SS: 37.1 -> 6.9 | Hurt |
| Increase p0 from 0.0 to 0.2 (ablation) | FUSS supervised and semi-supervised models · FUSS test set | SS: 11.1 -> 35.4 | Helped |
| Add zero-source loss (ablation) | Supervised separation model · WSJ0-2mix 2-source validation | SI-SNRi: 15.9 -> 14.3 | Hurt |
| Add zero-source loss (ablation) | Supervised separation model · FUSS validation, p0=0.0 | SS: 10.9 -> 12.0 | Helped |
| Add zero-source loss (ablation) | Supervised separation model · FUSS validation, p0=0.2 | MSi: 12.6 -> 12.3 | Hurt |
| Add zero-source loss (ablation) | Supervised separation model · FUSS validation, p0=0.2 | SS: 29.5 -> 34.7 | Helped |
| Add mixture consistency projection (ablation) | MixIT · WSJ0-2mix 2-source validation | SI-SNRi: +0.5 dB | Helped |
Closest papers by meaning
- UNSSOR: Unsupervised Neural Speech Separation by Leveraging Over-determined Training Mixtures (NeurIPS 2023 · similarity 0.89)
- Weakly-supervised Audio Separation via Bi-modal Semantic Similarity (ICLR 2024 · similarity 0.87)
- ZeroSep: Separate Anything in Audio with Zero Training (NeurIPS 2025 · similarity 0.85)
- Independency Adversarial Learning for Cross-Modal Sound Separation (AAAI 2024 · similarity 0.85)
- ArrayDPS: Unsupervised Blind Speech Separation with a Diffusion Prior (ICML 2025 · similarity 0.83)
- Separate and Reconstruct: Asymmetric Encoder-Decoder for Speech Separation (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.