settag

SetTag is a terminal-first analysis assistant for DJ music libraries. It uses MAEST for genre and optional Discogs-EffNet heads for mood/theme and instrument evidence, then lets you stage and verify metadata changes before writing.

The default experience is a Textual app. The same executable also provides a plain, non-TUI CLI for scripts, redirected output, saved plans, and CI.

Metadata hygiene is a separate, model-free workflow. It finds suspicious comments, web addresses, duplicate or empty text values, and generated encoder markers, then lets you review individual field-level removals before writing.

Supported files:

• MP3, AIFF, and WAV with ID3 metadata
• FLAC with Vorbis comments
• M4A, M4B, and MP4 with MP4 atoms

Install

uv tool install settag
settag models download
settag "/path/to/music"

Or with pipx: pipx install settag. SetTag needs Python 3.10–3.14.

Platform support is inherited from essentia-tensorflow, which ships only prebuilt binary wheels. There is no pure-Python fallback and no source build worth attempting, so a platform without a wheel cannot install SetTag at all:

Windows and Linux on ARM are not supported. Upstream has never published wheels for either, on any release.

The analysis backend is a large download — roughly 100 MB on macOS and 290 MB on Linux — and the models are fetched separately on top of that.

Genre is the default and the only model loaded unless tasks are requested explicitly:

settag models download --tasks genre,mood-theme,instrument
settag analyze "/path/to/music" --tasks genre,mood-theme,instrument

The interactive app reads its default tasks from ~/.config/settag/config.toml:

[analysis]
tasks = ["genre", "mood-theme", "instrument"]

The config file is optional. --tasks overrides it for one TUI run:

settag "/path/to/music" --tasks instrument
settag "/path/to/music" --tasks mood-theme,instrument
settag "/path/to/music" --tasks genre,mood-theme,instrument

Set SETTAG_CONFIG or pass --config /path/to/config.toml to use another config file. Task precedence is --tasks, then the config file, then the genre default. The config file is read only for options the flags left unset.

How much audio the genre model reads

The genre model, MAEST, embeds one 30-second patch at a time and dominates the run: 15.5s against EffNet's 1.2s on a 482-second track. --genre-sample chooses how many of those patches it reads. Mood/theme and instrument always read the whole track, because they are cheap and their taxonomies want whole-track averaging.

Measured against the full-track answer over 14 tracks, middle preserved the rolled-up conventional genre on 14/14 and spaced on 13/14, both with a rank correlation above 0.98 across all 519 labels. What moves is the crowded 0.1-0.25 tail, where the model is not confident anyway. Fewer patches also means less averaging, so scores come out more peaked.

[analysis]
genre_sample = "full"

Changing this changes the evidence configuration digest, so tracks analyzed under a different setting are correctly reported as stale and reanalyzed.

Models are downloaded once into ~/.cache/settag/models; --model-dir on models, analyze, and the app points at another directory, and models download --force replaces files that are already present. They are not bundled with this repository or its Python distributions. Downloads and installed files must match the SHA-256 digests pinned in SetTag's model catalogue before inference can start. Review Licensing before using the default models in a professional, business, or revenue-generating workflow.

Run the app

Give SetTag a track or a directory:

settag "/path/to/track.mp3"
settag "/path/to/music/library"

For metadata cleanup without analysis, open the independent hygiene review:

settag hygiene "/path/to/music/library"

No model files are loaded. In the main app, H switches to the same hygiene step after any running analysis has stopped. Use --no-tui to print findings without changing files.

In an interactive terminal this opens the Textual app as a full-width track list. Press I whenever you want to toggle the details panel. With details open, the library looks like this:

  ✓  Track                       File genre  Analysis
  ✓  Eli & Dani - What Do...     None        Never
     Robin Schulz - Sugar.mp3    House       Up to date · 2026-07-23

 Current file metadata
   Standard genre: None
   SetTag status: Never analyzed
   Last analyzed: Never

 Selected for analysis.
 The audio model has not been loaded.

SetTag first scans existing metadata only. Tracks with no analysis, incomplete metadata for any configured task, or a changed task model/config are selected by default. Up-to-date tracks remain visible but unselected. The active task families are shown in the library context line. Adjust the selection, then press R to load MAEST, EffNet, or both and analyze only the selected tracks visible in the current filter. Selections in other filtered views are not included.

Analysis runs serially in a background worker. The selected batch is fixed when the job starts, but the interface remains available for navigation, filtering, and inspection. Each completed track is persisted immediately and becomes available under V while the next track is still running.

The library keys are:

Cancellation is cooperative because Essentia/TensorFlow inference cannot be safely interrupted halfway through a track. Completed results remain available for review and are saved in the local workbench; unprocessed tracks remain selected for a later run. Cancellation never writes metadata.

Background execution keeps the interface interactive, but inference remains CPU-intensive; a warm laptop and active fans are expected during a large batch. SetTag deliberately analyzes one track at a time rather than multiplying that load with parallel track workers.

Press V as soon as the first track completes to open review. If you stay in the library, the app switches to review when the full batch finishes:

  ✓  Track                       File genre    Analysis          Suggested          Write plan
  ✓  Eli & Dani - What Do...     None → House  New · 2026-07-23  Progressive House  Evidence + Genre

 Standard file genre
   None → House (staged)
   Suggested roll-up: Progressive House → House

 Review candidates
   Score cutoff ≥ 0.10 · maximum 5
    1. Electronic---Progressive House  0.664
    2. Electronic---Techno             0.269
   18 additional ranked scores stored for importing apps.

The review keys are:

Newly analyzed tracks with SetTag changes are checked for writing by default. Press Space to check or uncheck the highlighted track. While background analysis continues, review, genre editing, plan saving, and writing operate on the completed snapshot only. The in-flight track cannot enter a write until its analysis plan is complete. When a newly analyzed track has no conventional genre, SetTag visibly stages the conservative standard-genre suggestion there by default. It never replaces a non-empty genre automatically. Use E to change the staged value or clear it to preserve the empty genre; the before → after value remains visible in the list and inspector. If no candidate clears the review cutoff, the genre remains empty.

The automatic default and the editor's Use suggestion action remove the Discogs parent prefix. For an explicit allowlist of House-family labels, they also roll the detailed child label up to the stable conventional genre House:

Electronic---Progressive House → House
Electronic---Tropical House    → House

The inspector shows this transformation. Other model children keep their direct name; SetTag deliberately does not infer a family from a suffix alone (Witch House remains Witch House). Detailed labels and scores are unchanged in the SetTag evidence. Use E to open the genre screen, where you can restore the suggestion after opting out or enter the exact value you want.

W runs a complete preflight and shows one batch confirmation. Write is the default focused action, so Enter confirms it; Esc returns to review. SetTag then runs preflight again, writes each file through its format-native adapter, and reopens it to verify the result. Analysis errors disable writing for the batch.

After a successful write, the app stays open. Written tracks leave the review batch and appear as up to date in the library. Any skipped review tracks remain available, so you can continue reviewing or return to the library for another analysis batch before pressing Q.

Local workbench and restart recovery

The app saves each completed analysis and staged genre edit to a small local SQLite workbench. If you quit before writing, opening the same track or directory opens in the library and shows the saved result as Ready · date. The model is not rerun. Press V to review saved results, or select a ready track in the library to reanalyze it deliberately.

The workbench is private application state, not portable music metadata:

• macOS: ~/Library/Application Support/settag/state.sqlite3
• Linux: ${XDG_DATA_HOME:-~/.local/share}/settag/state.sqlite3
• Windows: %LOCALAPPDATA%\settag\state.sqlite3 — the path SetTag would use, though Windows cannot install the analysis backend today (see Install)

Override it for a run with --state-db PATH, or globally with SETTAG_STATE_DB. For example:

settag "/path/to/music" --state-db "/path/to/settag-state.sqlite3"

A cached result is ready only while the source size and modification time, analysis model, evidence format, and observed standard genre still match. Changing --top or --score-cutoff does not require reanalysis; SetTag applies the new review policy to the stored evidence. Metadata already embedded in the audio file is authoritative and supersedes an obsolete workbench entry. Successfully written and verified entries are removed from the workbench. A track renamed or moved between runs is found again by its audio digest when exactly one scanned file matches, so its saved result follows it; an ambiguous match is reported as missing rather than guessed. Metadata created by the older cutoff-filtered contract needs one reanalysis to populate the new bounded evidence bundle.

The database is deliberately not an export format. Use S in review or analyze --plan when you want an explicit, portable JSONL plan that another command or machine can preview and apply.

Review defaults can be adjusted without changing the evidence written to files:

settag "/path/to/music" --top 5 --score-cutoff 0.10

--threshold remains an alias for --score-cutoff.

Metadata hygiene

Hygiene is intentionally separate from genre analysis:

scan tags → flag suspicious values → choose fields → confirm → clean and verify

The first release recognizes comment-like and generated metadata in each supported container:

A comment such as electronicfresh.com is suggested for removal because it contains a web address. An ordinary DJ note remains untouched. Encoder markers such as Lavf62.12.102, empty values, and exact duplicate values are also suggested. These rules create review suggestions, never automatic writes.

Every finding is checked independently. Space includes or excludes one field-level suggestion, A toggles all findings, I shows the exact before and after values, and W runs preflight and opens one confirmation. Only checked suggestions are written. Titles, artists, albums, artwork, genres, SetTag evidence, and unselected comments remain unchanged.

Hygiene writes use the same temporary-copy, reopen-and-verify, and journal path as analysis writes. settag undo therefore restores cleaned values too.

Plain CLI mode

When stdin or stdout is not a terminal, settag PATH automatically becomes a plain dry run. It never prompts or writes. Force that mode in a terminal with:

settag "/path/to/music" --no-tui

The named commands are also plain CLI commands:

# Which build stamps SETTAG_VERSION
settag --version

# Human-readable dry-run logs
settag analyze "/path/to/music"

# Complete machine-readable audit records
settag analyze "/path/to/music" --output analysis.jsonl

# Read current tags without loading the model
settag inspect "/path/to/track.mp3"

# Every field, but no ranked score lines: readable across a directory
settag inspect "/path/to/music" --no-scores

# Create, preview, and apply a durable plan
settag analyze "/path/to/music" --plan settag-plan.jsonl
settag preview settag-plan.jsonl
settag apply settag-plan.jsonl

For deliberate automation, apply --yes suppresses only the confirmation. It does not suppress source, metadata, or plan validation:

settag apply settag-plan.jsonl --yes

analyze never writes. A reviewed plan is the only route to disk, in the app or through apply, so every write in SetTag passes the same preflight and verification. Standard genre editing is available in the app or in a saved plan because it must be staged explicitly per track.

Logging

Normal INFO output is compact:

INFO [1/1] /path/to/track.mp3
INFO   file genre tag: House (unchanged)
INFO   SetTag genres: none -> Electronic---House score 0.765
INFO   dry run: SetTag analysis bundle would change (6 internal fields); nothing written

The complete 519-label record is written with --output or exposed through debug logging:

LOG_LEVEL=DEBUG settag analyze "/path/to/track.mp3"

LOG_LEVEL accepts DEBUG, INFO, WARNING, ERROR, or CRITICAL. Essentia's known one-time Abseil/MLIR startup messages are filtered separately because native TensorFlow code writes them before Python logging starts.

Metadata safety

SetTag-owned model evidence and a conventional file genre are separate:

Each task's label field holds bounded ranked evidence, and its _SCORES twin holds the same labels and order with the raw model scores as one compact JSON value. SetTag stores the top 60 results without a score cutoff so consumers such as SetPath can apply their own policy. Genre evidence is exclusively MAEST-derived; EffNet cannot populate SETTAG_GENRE or a conventional genre field. Task updates are independent, so an instrument-only run replaces instrument evidence while preserving valid genre and mood/theme evidence.

The full field list, the per-task vocabulary declaration consumers should key on, and the SETTAG_PROVENANCE record are specified once, in DESIGN.md.

The standard genre is not part of the SetTag namespace. For newly analyzed tracks where it is empty, the app stages the conservative standard-genre suggestion by default; the user can edit or opt out before confirming the write. Title, artist, album, artwork, duplicate fields, and metadata owned by other software are preserved. Comments and other hygiene fields are preserved unless the user explicitly checks their removal in the separate hygiene review.

Before any batch write, SetTag verifies for every included track:

1. the audio has not changed since analysis. The check uses a digest of the audio samples only, so another tool retagging the file in between does not block the write, while a re-encode, edit, or truncation does;
2. the observed conventional genre;
3. the detected native adapter;
4. every reconstructed SetTag change;
5. any explicitly staged standard genre change.

The audio digest is checked once more immediately before each file is written. After each write SetTag reopens the file and verifies both the evidence and the expected standard genre. A batch preflight is all-or-nothing, although native file writes cannot form one transaction across multiple files.

Undoing a write

Every verified write is journaled with the tag values it replaced, so a write can be reverted. Press U in the app, or use the CLI:

# What has SetTag written?
settag undo --list

# Preview reverting the most recent write
settag undo --dry-run

# Revert the most recent write, or a named one
settag undo
settag undo 20260725T110349-8993c143

# Show more history, or revert without the confirmation prompt
settag undo --list --limit 25
settag undo --yes

The journal lives in its own database, separate from the workbench cache, so clearing the cache never destroys undo history:

Undo restores exactly what a write changed: the SetTag-owned fields listed above, the conventional genre tag when that write explicitly staged an edit, and any fields explicitly removed by a hygiene write. A track that had no SetTag metadata beforehand is returned to having none rather than being left with debris.

Some limits worth knowing:

• It restores tag values, not bytes. mutagen rewrites the tag block on save, so the file will not regain its pre-write SHA-256 even after a perfect undo.
• It is per write, not "restore original". Two writes to one file produce two journal entries; undoing the newest returns the state after the first write.
• It leaves other software alone. Edits made elsewhere between the write and the undo are untouched.
• It refuses changed files. If a file was modified after SetTag wrote it, that file is skipped and named; --force overrides.

Entries older than 90 days are pruned.

Saved plans

A plan is one JSONL record per track on the settag.plan/v5 schema. It carries the bounded evidence separately from SetTag's current review selection, the observed file genre separately from an optional staged target, the audio digest preflight checks against, and the human-readable change lines preview prints. The record layout is specified once, in DESIGN.md.

New plans also carry source.owned_sha256, a digest of the exact observed SetTag-owned metadata. Preflight rejects changes to those values even when the display summary is identical. Older plans without this digest require the whole file to remain unchanged, including unrelated tags.

Plans produced by analyze --plan set target_file_genre to null. Plans saved from the app retain explicit edits. preview is the built-in human-readable renderer; users do not need jq.

Every apply performs preflight both before and after confirmation. A source change or an analysis-error record rejects the whole plan before writing.

Plans on the pre-release settag.plan/v4 schema still apply. They predate the audio digest, so preflight falls back to comparing the whole file for them: a tag write by another tool blocks a v4 plan where a v5 plan would proceed. Any other schema is rejected with an explicit error.

Scores and models

The pinned production models are:

• MAEST genre:
  • discogs-maest-30s-pw-519l-2
  • genre_discogs519-discogs-maest-30s-pw-519l-1
• optional Discogs-EffNet metadata:
  • discogs-effnet-bs64-1
  • mtg_jamendo_moodtheme-discogs-effnet-1
  • mtg_jamendo_instrument-discogs-effnet-1

When genre and EffNet tasks are requested together, one 16 kHz audio decode feeds both model stacks. Mood/theme and instrument always reuse one EffNet embedding pass.

Each score is the mean class-wise sigmoid activation across analyzed audio patches. It is useful for ranking and applying a score cutoff, but it is not demonstrated to be calibrated confidence or probability. Scores do not need to sum to 1.

The review cutoff controls review markings and suggestions only. It does not remove scores from the bounded evidence bundle:

settag analyze "/path/to/music" --top 5 --score-cutoff 0.10

Licensing

SetTag source is licensed under AGPL-3.0-only. The default inference workflow also uses separately licensed components and model files:

• essentia-tensorflow is AGPL-3.0-only; UPF also offers proprietary licensing
• TensorFlow runtime is Apache-2.0
• UPF publicly offers the Essentia model weights and metadata for non-commercial use, with proprietary licensing available

UPF's public documentation currently identifies the exact Creative Commons variant inconsistently as CC BY-NC-ND 4.0 and CC BY-NC-SA 4.0. The model repository's licence file is also internally inconsistent, and the pinned model metadata does not specify a licence. Both stated variants restrict the public grant to non-commercial use. Until UPF provides model-specific clarification, SetTag does not assert permission to redistribute or publish adapted model files.

Personal, educational, or research use may fall within the public model terms. Professional, business, or other revenue-generating use is not clearly permitted and may require separate permission from UPF or a different analysis backend. Downloading the models separately does not change their terms.

Metadata produced by SetTag can be exported to and imported by compatible music-library applications. Users remain responsible for ensuring that their selected analysis backend permits their intended use.

Exact attributions and dependency notices are in THIRD_PARTY_NOTICES.md.

See:

• https://essentia.upf.edu/licensing_information.html
• https://essentia.upf.edu/models.html
• https://essentia.upf.edu/models/LICENSE

Development

uv sync --group dev
uv run pytest
uv run ruff check .
uv run ruff format --check .
uv run ty check
uv build

GitHub Actions runs the same lint, format, type, and test checks on every supported Python version (3.10 through 3.14) for each push to main, pull request, and release tag.

The default test run uses Textual's headless app runner and synthetic audio; it does not run model inference. A small set of smoke tests against the real genre model is opted into separately and skips itself when the models are not downloaded:

uv run pytest -m models
uv run settag models status --tasks genre,mood-theme,instrument
uv run settag analyze "/path/to/track.flac" --tasks genre,mood-theme,instrument

Architecture and safety contracts are documented in DESIGN.md.

Optional audio evidence for ranking experiments

SetTag can export pooled EffNet embeddings without writing them into your music files. This is optional evidence for consumers such as SetPath, which must remain usable without SetTag or its models. It does not declare energy, vocal prominence, or any other calibrated perceptual axis.

settag analyze "/path/to/music" --tasks mood-theme,instrument --embeddings embeddings.jsonl

The export reuses the same EffNet embedding pass as those tasks. Genre-only runs cannot export it; no extra model is silently enabled. The output must be a new file. Completed tracks are flushed individually, and any track failures produce a nonzero exit status and diagnostics on stderr. The file contains successful records only; an interrupted or partially failed batch is not complete coverage.

Each audio-embedding/v1 record carries source identity, model weights digest, output tensor, preprocessing, mean pooling, L2 normalization, dimensions, patch count, and the vector. The current EffNet model produces 1,280 dimensions. These are whole-track embeddings, not intro/outro or phrase measurements.

A consumer must compare compatible feature spaces and treat missing evidence as unknown. Different producers can implement the same contract; there is no runtime connection between SetTag and SetPath. SetPath currently consumes these exports only through its explicit offline evaluation commands. Export from the final file state: its importer verifies the whole-file SHA-256 and rejects files retagged or replaced since export.