TrackAudioFeatures

Tempo in beats per minute, rounded to the nearest integer. The detector is a dynamic programming beat tracker computed over the onset-strength envelope.

Where it comes from

librosa.beat.beat_track(onset_envelope=...). Estimates in the mid range (90–180 BPM) tend to be solid; half-time genres can double/halve (e.g. a 75 BPM track reported as 150).

How it is used

• Filter in /tracks/{id}/similar (bpmMin, bpmMax).
• Shown in ManageAlbumPageClient (Studio) in the analysis section.
• Tempo heuristic (slow / mid / fast) in the TrackInsightsModal.

How it could be used

• Building BPM-compatible DJ playlists (beatmatching).
• Workout / running playlists by cadence band.
• "Similar tracks within ±5 BPM" recommendations.

Confidence of the BPM estimate, derived from the variance of the intervals between detected beats. Below ~0.5 the estimate is considered unreliable and the UI shows a warning.

How it is used

• The UI shows BPM with a confidence %.
• Recommendation pipelines may ignore low-confidence BPM.

Detected tonal note (tonic). Chromagram via librosa.feature.chroma_cqt(), compared against Krumhansl-Kessler profiles for each of the 12 pitch classes.

How it is used

• Shown in the track detail UI.
• Together with Mode, displays the tonality ("A minor").

How it could be used

• Harmonic mixing (Camelot / Open Key wheel) — find compatible tracks for transitions.
• Suggestions for covers/remixes in the same key.

Tonality mode (major or minor). Detected as the higher correlation between the mean chromagram and the major/minor Krumhansl-Kessler profiles.

Distance between the best tonal profile and the second best. Atonal or very percussive tracks tend to have low confidence.

Integrated loudness in LUFS (Loudness Units relative to Full Scale). Computed by pyloudnorm following the international broadcast standard. Typical targets:

• Spotify / Apple Music / YouTube: −14 LUFS
• Tidal: −14 LUFS
• Broadcast (EBU R128): −23 LUFS

How it is used

• LoudnessMeter in the TrackInsightsModal — shows LUFS vs target.
• Classified as "high / ok / low" via loudnessInfo() (insights.ts).
• Shown as text in the Studio analysis section.

How it could be used

• Automatic loudness normalisation in the player (applied gain).
• Warning artists that their master is "below the streaming target".

Sample peak in dB Full Scale. Indicates whether the master is digitally clipping (a value > 0 dBFS is impossible with integer samples; it can occur in float).

True peak, reconstructed via polyphase oversampling (4×). Lets you detect inter-sample peaks — peaks that appear between samples after D/A reconstruction and can cause clipping in lossy codecs.

Amplitude envelope reduced for visualisation. 800 RMS magnitude samples per window, normalised to 0–1. The front-end (Waveform) downsamples to the canvas size via downsamplePeaks().

Where it comes from

When the audiowaveform (BBC) binary is available it is used for fast extraction; otherwise numpy computes windowed RMS.

How it is used

• Rendering the waveform in the player (Studio / Web).
• Visual scrubbing during playback.

Chromaprint acoustic fingerprint — a compact string representing the spectral content of the recording. Robust against re-encoding and small pitch/tempo variations.

Where it comes from

The external fpcalc binary (from the AcoustID project), invoked by the Python sidecar.

How it is / could be used

• Duplicate detection — compare fingerprints with tolerance to identify uploads of the same audio under different names.
• AcoustID lookup — feed the AcoustIdMatch column (pending).
• Detecting re-uploads after a takedown.

Best AcoustID recording GUID found from the Fingerprint. Currently always null — the integration with the public AcoustID API is reserved but not active yet.

When to enable

• Cross-link with MusicBrainz (canonical metadata, ISRC, releases).
• Detecting tracks that already exist on other platforms.

Estimated distribution between speech, music and noise across the file. A heuristic combining RMS, spectral flatness and zero-crossing rate.

How it is / could be used

• Detecting podcasts/spoken-word incorrectly published as music.
• Filtering spoken previews/intros.
• Smart skipping ("skip talking parts").

Timestamp marking the end of acoustic analysis. Used to invalidate data after algorithm changes (re-analyse when stale).

Probability that the track is "danceable" (class 0 of the binary classifier). Combines pulse, rhythmic regularity and spectral stability.

How it is used

• An axis of the AudioDnaRadar (TrackInsightsModal).
• Filters and ranking in dance/club playlists.

Perceptual energy derived from RMS via a logarithmic mapping. Not an ML model — it is a direct computation on the waveform.

How it is / could be used

• An axis of the AudioDnaRadar.
• energyMin filters in /tracks/{id}/similar.
• Curating "high-energy" vs "chill" playlists.
• Ordering tracks in a set/album along an emotional arc.

How much the loudness varies over time. High values indicate dynamic material (classical, jazz); low values indicate compressed/loud material (commercial pop, EDM).

How it could be used

• Flagging over-compressed masters.
• Contributing to the track's "DNA" radar.
• Recommending to listeners who prefer dynamic material.

Probability of the "voice" class (class 1) — how much vocal content is in the track. Independent of the editorial IsInstrumental flag: it can be used to validate the artist's declaration.

How it is / could be used

• An axis of the DNA radar.
• "Karaoke / Instrumental" filters.
• Cross-checking Track.IsInstrumental to surface upload inconsistencies.

Probability that the track conveys a "happy" mood (class 0 of the binary classifier). A "happy vs not happy" classifier trained on MTG-Jamendo labels.

How it is used

• Feeds the Valence heuristic.
• Directly in the UI for "Feel Good" playlists.

Probability of a "sad" mood. A component of the Valence calculation.

Probability of an "aggressive" mood (typical of metal, hard rock, heavy rap). A component of Arousal.

Probability of a "relaxed" mood. A component of Arousal (inverted).

Probability of a "party feel" — uplift, energy, danceable BPM.

Multi-label probabilities for 56 mood/theme tags from the MTG-Jamendo dataset. Each tag has an independent score in [0,1]. Example: { "energetic": 0.82, "epic": 0.61, "dark": 0.04, ... }.

Available tags

action · adventure · advertising · ambiental · background · ballad · calm · children · christmas · commercial · cool · corporate · dark · deep · documentary · drama · dramatic · dream · emotional · energetic · epic · fast · film · fun · funny · game · groovy · happy · heavy · holiday · hopeful · inspiring · love · meditative · melancholic · melodic · motivational · movie · nature · party · positive · powerful · relaxing · retro · romantic · sad · sexy · slow · soft · soundscape · space · sport · summer · trailer · travel · upbeat · uplifting

How it could be used

• Sync licensing: matching tracks to briefs ("epic / dramatic / trailer").
• Auto-tagging editorial moods.
• Contextual recommendation: "workout music" → energetic + upbeat + powerful.

Mean spectral energy as 8 ERB-band values computed at 44100 Hz. For display these are grouped into four bands:

• sub: ~20–80 Hz (kick, sub-bass)
• bass: ~80–250 Hz (low end, electric bass)
• mid: ~250–4000 Hz (vocals, instruments)
• high: ~4–20 kHz (cymbals, air, brightness)

How it is used

• EnergyBands chart in the TrackInsightsModal.
• Visualising the master's "sonic profile".

How it could be used

• Detecting masters lacking bass or with excessive brightness.
• Recommending by the "colour" of the mix (warm vs bright).

The arousal axis of the valence/arousal plane (Russell's circumplex model). How intense and active the track is vs calm and contemplative.

How it is used

• The Y axis of the MoodMap (Studio).
• An axis of the AudioDnaRadar.

The valence axis — emotional positive vs negative. Values near 1 = happy/uplifting; near 0 = sad/melancholic.

How it is used

• The X axis of the MoodMap.
• valenceMin / valenceMax filters in /tracks/{id}/similar.
• An axis of the DNA radar.

A 1280-dimension embedding extracted from the penultimate layer of the discogs-effnet-bs64-1.pb model, with mean pooling over time. Captures musical semantics (style, instrumentation, sonic texture).

How it is used

-- /tracks/{id}/similar — approximate cosine search
SELECT t."Id",
       1 - (taf."EmbeddingDiscogs" <=> '[…]'::vector(1280)) AS similarity
FROM tracks t
JOIN track_audio_features taf ON taf."TrackId" = t."Id"
WHERE taf."EmbeddingDiscogs" IS NOT NULL
ORDER BY taf."EmbeddingDiscogs" <=> '[…]'::vector(1280)
LIMIT 50;

How it could be used

• "More like this" anywhere in the app.
• Personal radio blending embeddings of liked tracks.
• Catalog clustering for discovering emerging genres.
• Outlier detection within an album.

Drives the retries of EnrichTrackAudioConsumer. Initial value Pending at the end of Phase 1.

Timestamp of the last successful enrichment.

Attempt counter. The worker increments it on each run. After the configured maximum (3), it records EnrichmentStatus = Failed and releases the track to Ready (fail-open).

Last error message. Useful for diagnostics in Studio (admin UI) and for alerting/observability.