MUSIC PROFESSIONAL - Stem Source Separation Expert - Proof of Experience Required

Current open-source source separation models (HTDemucs, Open-Unmix, MDX-Net, RS-RoFormer family) separate mixtures into only 4–6 stems. Real music mixtures typically contain dozens of distinct instruments, so this coarse granularity limits downstream applications like remixing, music education, transcription, and music information retrieval. This project collects annotations to extend SOTA source separation models beyond the standard 4-stem paradigm by having annotators identify specific instruments and instrument families present in short audio segments.