Docs Analysis
Analysis
One pass per track, cached on your disk. Everything the planner and the renderer do later depends on what comes out of it.
What a pass produces
- Grid. Every beat and downbeat, plus the 4-bar phrase grid you count in
- Tempo. BPM, and how steady it stays. A track whose grid cannot be trusted is held back rather than used
- Key. Key and how confident we are, in Camelot wheel notation
- Level. Loudness, so no track arrives louder or quieter than the one it follows
- Energy. An energy curve across the whole track, second by second
- Shape. Labelled sections: intro, verse, breakdown, build, drop, chorus, outro
- Per section. How much bass, vocal and drum each section carries, and how busy it is
- Timbre. A fingerprint of the track's texture, so you can look for something that sounds like this
- Stems. Drums, bass, vocals and melody, pulled apart so a blend can hand over one at a time
- Flags. Whether it has vocals, whether the tempo holds, whether the grid can be trusted
The grid, and how it is checked
Beats and downbeats come from a neural tracker. On top of the tracker we grade how tightly the resulting grid sits on the actual kick, per track, which is a property no public benchmark asks for and the one that decides whether a technically correct mix feels right. The median across our own library is 8.5 milliseconds.
Phrases matter as much as beats. A 4-bar phrase grid is derived from the downbeats, and seams are anchored to phrase boundaries rather than arbitrary positions, which is why joins land where a listener expects a change.
Quarantine
When tempo stability comes back at or near zero, the grid is refused and the track is quarantined. It stays out of planning and out of the measured map. This is deliberate: a set locked to one tempo will warp every track to that tempo, and warping a track whose grid is wrong produces exactly the artefact that ends a night early.
Rescuing quarantined tracks with alternative trackers is on the roadmap. Where a rescue fails, the track stays quarantined and is listed as needing a manual grid, rather than being silently promoted.
Key, and what key is for
Key is detected from the audio, with a confidence value, and reported in Camelot notation because that is what harmonic mixing uses. Key is one term in candidate scoring, not a veto: a strong pair with an awkward key relationship can still win, and a perfect Camelot neighbour with a bad rendered seam still loses.
Accuracy
The analysis pass is the part of the product we hold to public benchmarks, on held-out data, and publish even when a number is against us. The board carries key, beats, downbeats, tempo and structure, each with its dataset and split.
Structure detection was rebuilt in August 2026. Measured at the operating point and decode rate the application runs, it scores coarse HR.5F 0.1966 / HR3F 0.4025 held-out (n=153): level with our own Foote reproduction at the strict tolerance and 0.070 behind it at the lenient one. The denser 0.2341 / 0.4823 is an ablation, not the product. It reliably finds blend anchors, which is what the engine uses it for. See the structure row on the board for the numbers and the caveats.
Re-analysis
Adding a new kind of analysis means re-reading your files, which takes real time on a large library. So analysis changes get bundled deliberately rather than shipped one at a time. Playback-time changes, including everything about how a seam is built, need no re-analysis at all.