First fully supervised recipe on real sample pairs pushes music sample identification past SOTA
serrjoa · x · 2026-09-21
Researchers from MTG and Sony present SI Embeddings (SIE) in a new arXiv paper, the first fully supervised training recipe for music sample identification using real sample pairs.
- Prior work relied on artificially constructed pairs due to lack of annotations; the old SOTA only partially generalizes to real pairs
- SIE achieves state-of-the-art results on three benchmarks, including a large-scale test set
- The authors show architecture and training recipe contribute substantially, not just the real data
- The recipe establishes a strong foundation for future SI research
More from Multimodal
- VoiceStudio: open-source local ElevenLabs alternative with voice cloning in 646 languages — thione · 2026-09-21
- GPT Image 2.5 water-level photo prompt shared in full for realistic tourist shots — techhalla · 2026-09-21
- TripoAI launches Smart UV: AI-driven seam drawing and unwrapping in 5 seconds — majidmanzarpour · 2026-09-21
- One-word Midjourney prompts: generating art from the archaic word 'Sooth' — tisch_eins · 2026-09-21
- Traditional 3D blended with Jax diffusion frame-by-frame animation — makeitrad1 · 2026-09-21
- Higgsfield's Open API offers pay-per-generation access to 50+ models — SolidSnakesCranny · 2026-09-21