superset
liveSuperset automates the part of set preparation that doesn't require taste: curating a coherent set from a library of thousands of tracks. It applies the same audio feature extraction that powers music recommendation systems to DJ curation.
The pipeline runs locally on Apple Silicon. Rekordbox XML is parsed to extract a track catalog, then librosa and Essentia compute 40+ features per track — BPM, key, spectral centroid, timbral texture, energy envelope, and more. KMeans and DBSCAN cluster the library into vibe and energy groups, and a greedy graph traversal generates a setlist that respects harmonic mixing rules and energy arc.
The output is a rekordbox-compatible XML playlist — it drops directly into the existing DJ workflow, no intermediate steps. A CustomTkinter GUI allows interactive library exploration and manual override at any point in the process.
// highlights
40+ spectral, timbral, and rhythmic features per track via librosa + Essentia
Runs entirely on-device — no API calls, no cloud dependency
KMeans + DBSCAN dual-clustering for energy and vibe segmentation
Exports rekordbox-compatible XML playlists
CustomTkinter GUI with real-time library exploration
Genre-specific parameter modes (house, dubstep)
Optimized for Apple Silicon via MLX