Training competitive music generators and recommendation engines requires structured, precise audio annotations made by real musicians. We supply human-annotated, multi-dimensional musicology datasets — every record reviewed and QA-certified before export.
Every track is analyzed locally using Essentia.js WASM extractors. Key and BPM are extracted directly from the PCM audio buffer and verified against the annotation before the record is accepted.
A trained musicologist listens to each track and manually annotates every section: instrumentation, melodic contour, vibe, narrative description, and production characteristics. No AI generates the annotations — every judgment is human.
Every annotation is reviewed against strict quality constraints before approval: no cross-section references, complete instrument lists, validated vibe tags, and mandatory production assessment. Approved records are exported as structured JSON.
Every timeline block contains a strict 4-sentence structure describing: 1) Foreground melody/vocals, 2) Middle-ground keys/rhythm, 3) Background bass/drums, and 4) Dynamic energy shifts.
Detailed semantic tracking of panning distributions, sidechaining, effects (reverb/delay wetness), harmonic saturation, EQ filtering, and overall fidelity standards (studio vs. Lo-Fi).
Vocals and instruments are mapped using formal contours (undulating, conjunct, disjunct, static, ascending/descending arches) allowing generative models to train on target melodic curves.
Strict QA checks guarantee 0:00 alignment, prevent cross-reference shortcuts ("same as", "repeats"), mandate specific outro closures, and verify instrument presence per-block.
Loading database records...