Bombista

You have a recording, and you have the text of what's said or sung in it. Bombista works out when each line happens — and tells you which lines it isn't sure about, so you check three instead of proofing forty.

Heads up: the first run downloads the faster-whisper medium model (~1.4 GB) and caches it locally. Every run after that is offline.

Bombista is not an automation tool. It is a triage tool.

Forced aligners already exist — aeneas, the Montreal Forced Aligner, whisperX — and they all hand you timings with no opinion about them. So you either trust the lot or re-check the lot, and on a deadline you re-check the lot.

What Bombista adds is the review loop:

1. Per-line confidence bands — HIGH / REVIEW / FAIL, each with a named reason: clean-anchor, ambiguous, lead-fallback, uncorroborated, gap-outlier, no-anchor, override.
2. A report that says which lines to check and why — as markdown, as JSON, or as a self-contained HTML page with a play button per line that seeks the audio to that line's onset.
3. A correction pass that re-runs in well under a second, because the transcription is cached.

The claim that survives contact with a real deadline: timings, plus triage, plus instant correction.

Who it's for

Nothing here is music-specific:

• anyone subtitling video
• lyric-video and karaoke makers
• accessibility captioning
• educators building synced read-along texts
• audiobook ↔ ebook sync
• theatre surtitles — this exact job, done today by opera houses with worse tools

Runs offline

No API keys, no GPU, no network. faster-whisper medium (~1.4 GB, cached locally on first use), CPU int8, roughly 50 s for a three-minute song on an M-series Mac.

This is a property worth protecting, and it is why Bombista times but does not translate. Turning one language into another needs a model with opinions about meaning; working out when a line happens does not. Structural conversion — plain text into the canonical JSON shape — is deterministic and stays in the tool. Linguistic translation is your business, with your own tools.

---

Install

Bombista needs Python 3.11 or newer and expects to live in a virtual environment.

git clone https://github.com/jorgevallejos/bombista.git
cd bombista
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

The venv is not optional, and it is not relocatable. Console-script shebangs bake in an absolute interpreter path, so moving or renaming the checkout after installing will break the bombista command with a bare exit 127. If that happens, delete .venv and repeat the three commands above — nothing else is lost.

Verify:

bombista --help
python -m pytest

If bombista is not on your PATH — an unactivated venv, a pip install --user — every command below also works as python -m bombista.cli.

---

A worked example

You have a recording and its lyrics. You want to know when each line lands.

The song below — Río de Sal, twenty lines — is invented, and so are its numbers; they are representative of a real run, not a transcript of one.

1. Align

bombista align audio/rio-de-sal.m4a lyrics/rio-de-sal.json -o staging/rio-de-sal --lang es

The lyrics input can be a CP song JSON or a plain text file, one line per line. In a text file, blank lines and [Bracketed] section labels are stripped — and reported, so the removal is visible rather than silent.

About 50 seconds later, one line of output:

HIGH 18 / REVIEW 2 / FAIL 0 — timeline: staging/rio-de-sal/rio-de-sal-timeline.json — report: staging/rio-de-sal/rio-de-sal-qa-report.md — words: staging/rio-de-sal/asr-words.jsonl

Eighteen lines Bombista is confident about. Two to look at. That number is the product.

2. Read the report

The QA report opens with the provenance of the run — which audio file, its sha256, the model, the language — then gets to the point:

## Needs attention

| line | band | canonical text | ASR context | start | end | dur | signals |
|------|------|----------------|-------------|-------|-----|-----|---------|
| 12 | REVIEW | No soy la orilla, soy la sed, | No soy la orilla soy la se No soy el mar que | 61.35 | 64.88 | 3.53 | ambiguous |
| 19 | REVIEW | Vuelvo a empezar. | Vuelvo a empezar Contando faros en la niebla | 92.40 | 111.75 | 19.35 | lead-fallback |

- Line 12: re-run `align` with `--anchor 12=<seconds>` and `--words staging/rio-de-sal/asr-words.jsonl` to skip re-transcription (candidate start was 61.35 s).

ambiguous means the transcription heard something close to two different places. lead-fallback means the line's end had to fall back to the last transcribed word — which is why line 19 claims to run for 19 seconds.

To judge those two lines you need to hear them. Add --emit html:

bombista align audio/rio-de-sal.m4a lyrics/rio-de-sal.json -o staging/rio-de-sal \
  --lang es --emit html --words staging/rio-de-sal/asr-words.jsonl

That writes rio-de-sal-review.html — the same report as a page you open in a browser, with a play button on every row that seeks the audio to that line's onset, the REVIEW rows highlighted, and the fix command pre-written beside each one, click to copy. No server, no network, no build step.

3. Correct

Line 19 actually starts at 91.2 s. Pass it in, and reuse the cached transcription:

bombista align audio/rio-de-sal.m4a lyrics/rio-de-sal.json -o staging/rio-de-sal-fixed \
  --lang es --words staging/rio-de-sal/asr-words.jsonl --anchor 19=91.2

0.07 seconds. The expensive part was the transcription, and that is on disk. Correcting is free, so you correct until it is right instead of settling.

4. Apply

bombista promote staging/rio-de-sal-fixed/rio-de-sal-timeline.json lyrics/rio-de-sal.json

promote backs the file up next to itself, refuses on a line-count mismatch, warns loudly if the lyrics have changed since alignment, replaces only the timeline, and prints a per-line diff.

Total: about four minutes, of which roughly ninety seconds is human attention on two lines out of twenty.

---

Other outputs

--emit is repeatable, and passing it replaces the default rather than adding to it:

--emit	what you get
timeline (default)	the native envelope: { timelineVersion, leadIn, timeline }
songjson	the full canonical song JSON with the timeline merged in
report-json	the QA report as data — bands, signals, provenance, per-line
srt	one subtitle file per language present
lrc	one lyric file per language present
html	the offline review page described above

asr-words.jsonl and the markdown QA report are always written, whatever you ask for.

The timeline

Bombista's native artifact is a timeline: an ordered list of {start, end} spans, matched to lines by position, containing no words.

{
  "timelineVersion": 2,
  "leadIn": { "durationSec": 7.26, "source": "measured", "confidence": "low", "apply": false },
  "timeline": [
    { "start": 0.00,  "end": 5.84 },
    { "start": 5.84,  "end": 9.64 }
  ]
}

Two properties are worth knowing about.

It is language-independent. Retranslate every line into Dutch and the timings still hold. The words change; when they land does not.

Entry 0 always starts at 0.00. The silence before the first word is banked separately in leadIn rather than folded into every timestamp. This is a correctness measure, not tidiness: the measured lead-in is the single least reliable number the tool produces — ASR has a known habit of clamping the first word toward zero — and isolating it keeps one uncertain value in one hand-editable field instead of contaminating all twenty.

Whether to apply the lead-in is the consumer's decision, never Bombista's. Bombista always measures, always normalises, always records.

The cost of matching lines by position is positional fragility: insert one line into the lyrics and every timestamp after it is silently wrong. promote guards against this with a hash over the lines, and warns when they no longer match.

Commands

bombista align AUDIO SONG_JSON_OR_LYRICS_TXT -o STAGING_DIR
    [--lang es] [--model-size medium] [--anchor LINE=SECONDS]
    [--words STAGING/asr-words.jsonl] [--emit timeline|songjson|report-json|srt|lrc|html]

bombista promote STAGING/SONG-timeline.json SONG.json

bombista migrate SONG.json [--dry-run]

extract is a working alias for align — the original verb, kept so old commands still paste. migrate is a one-off for timelines produced before the leadIn model existed.

Development

Strict red → green → refactor; tests green before every commit.

python -m pytest

All fast except one integration test that runs the tiny Whisper model against a committed 12-second fixture.

License

MIT — see LICENSE.

---

Bombista is part of Tramoya, the stage machinery behind Chango Pepper. A bombista is the player of the bombo legüero, the drum that sets the pace for the ensemble. The repository was called timeline-extractor until August 2026.