Changelog

API version history, new features, breaking changes, and deprecations. Subscribe to updates at hello@brizm.dev.

v2.0.1 — Measurement accuracy (2026-09-10)

Four corrections in the shared analysis engine. No endpoint, field or schema changes — only values. If you pin analysis values in tests or compare them against results computed earlier, read the shifts below.

Fixed

• Tempo octave errors. tempo (and bpm on /v1/bpm and /v1/key-bpm) could report double the real tempo — 240 where the track runs at 120 — on material where the beat model locks onto eighth notes rather than the quarter-note pulse. The octave correction now always runs. Tracks without a subdivided pulse are unaffected: our four reference tracks return byte-identical values.
• The beat grid follows the corrected tempo. On those same tracks GET /v1/beatgrid/{id} returned beats at the subdivided rate — a 12-second clip at 120 BPM came back as 47 beats spaced 0.25 s apart. beats and downbeats now sit on the corrected tempo, so on affected tracks the beat count halves, downbeat spacing doubles, and the grid's tempo agrees with /v1/bpm. Genuine double-time material reports the half-time grid, matching the tempo value it always returned.
• True peak never reads below sample peak. true_peak_dbtp could come back up to 0.21 dB below the file's own sample peak when the loudest point was a single sample — drums and other transient material. True peak is now floored at the sample peak, as ITU-R BS.1770-4 requires, so those readings rise by up to 0.21 dB. Sustained material is unchanged. Mastering output limited to a true-peak ceiling no longer overshoots it by the same amount.
• Async results carrying non-finite numbers. A job whose result contained a non-finite number — the integrated loudness of a silent file, for one — produced a callback body that was not valid JSON and could not be delivered, after the analysis had already succeeded. Those values are now delivered as null.

Changed

• Loudness is EBU Tech 3341 compliant. Released 2026-07-22; recorded here late. The K-weighting pre-filter now uses the exact ITU-R BS.1770-4 design, and the meter passes EBU Tech 3341 cases 1–3 and Tech 3342 within tolerance. loudness, loudness_lufs and lra read slightly lower: broadband music by at most 0.05 LU, tonal or mid-heavy material by up to about 0.29 LU, while low-frequency-dominant tonal content reads up to 0.03 LU higher. Use a 0.3 LU tolerance when comparing values computed on either side of that date.

Note on cached results

• Analyses already stored in our cache keep the values they were computed with; this release changes new analyses. Contact hello@brizm.dev if you need specific tracks recomputed.

v2.0 — Route atomicity (2026-04-17)

Changed

• Replaced legacy /v1/compat/spotify/audio-features* routes with canonical /v1/audio-features* + ?format=spotify opt-out. The Spotify-shaped response is now served from the canonical URL via a query parameter instead of a parallel route.
• Replaced legacy /v1/analyze* routes with canonical /v1/audio-features*. Handlers are unchanged — purely a URL rename per Brizm's "leader not shim" URL discipline.

v1.2.0 — 2026-04-09

Added

• _meta envelope on all responses — cache status, latency, credits charged, credits remaining, and a trace ID for support tickets.
• GET /v1/compat/spotify/audio-features/{id} — Spotify drop-in compatibility shim. Same response schema, same field names, same scales.
• Song structure endpoint: POST /v1/structure — detects intro, verse, chorus, drop, breakdown, and outro segments with timestamps.

Changed

• GET /v1/analyze now returns integer 0–100 scales for mood dimensions (previously 0.0–1.0 floats). Native clients should migrate. The Spotify compat endpoint still returns floats to preserve Spotify's schema.

Fixed

• Rare race condition in credit deduction that could double-charge on retry when a client retried before the first response was acknowledged.

v1.1.0 — 2026-04-05

Added

• Mood heads trained on 114K tracks. New dimensions: energy, danceability, happiness, acousticness, instrumentalness, speechiness.
• GET /v1/embedding/{id} — raw 768-dimensional MAEST embeddings (Pro tier and above).
• GET /v1/similar/{id} — hybrid similarity search via pgvector, combining embedding distance with tempo and key compatibility (Pro tier and above).

Changed

• Removed librosa dependency — 100% standalone DSP. Roughly 40% faster cold start and tighter memory footprint on bare metal.

Fixed

• Key detection confidence was systematically underestimated on short tracks (<60 s). Recalibrated against the GiantSteps dataset.

v1.0.5 — 2026-03-27

Added

• On-demand resolution cascade — cache misses now trigger lazy compute via Deezer 30-second previews, so first-ever lookups succeed instead of returning 404.
• Negative cache (not_found table, 24-hour TTL) — skips re-hitting paid APIs for tracks we've already failed to resolve.

Changed

• Bare metal FastAPI now backs all sync endpoints directly, with Modal GPU overflow kicking in only on 503 from bare metal.

Fixed

• Chromaprint-based lookups now handle 96 kHz audio correctly — previously returned false negatives on high-sample-rate uploads.

v1.0.0 — 2026-03-15

Added

• Initial public launch.
• Sync endpoints: GET /v1/analyze, GET /v1/key-bpm, GET /v1/resolve/{id}, GET /v1/beatgrid/{id}, POST /v1/bpm, POST /v1/key, POST /v1/lufs, POST /v1/analyze/upload.
• Async endpoints (202 → poll or webhook): POST /v1/separate, POST /v1/transcribe, POST /v1/master.
• Five-layer D1 cache: file hash, chromaprint, ISRC, Apple ID, normalized metadata.
• Credit-based billing with Free, Starter, Pro, Scale, and Enterprise tiers.
• Test (tl_test_*) and live (tl_live_*) API keys.
• 9.3M track identity database for instant cross-platform resolution.