Reel: Caption and Transcript Governance

Caption and transcript governance for the video and audio on a Statamic site. Reel does not generate captions. It finds the caption file you already have, reads it the way a player and a reader will, says what is wrong with it, lets an editor fix it, and refuses to publish an entry whose media has no captions or transcript at all. Pair it with whatever caption generator you already use; the point of Reel is what happens after that tool has had its turn, because uncorrected automatic captions do not meet WCAG 1.2.2.

Reel is a paid addon for Statamic 6 on PHP 8.2 or newer.

What it does

For every video and audio asset in your containers, Reel reports where it stands on WCAG 2.1 success criteria 1.2.1 to 1.2.5:

A criterion Reel cannot decide is reported as Review, never as a pass. An audio-description flag nobody has set is "not reviewed", not "needed" and not "provided".

The caption file itself gets three kinds of attention:

• Hard errors, which a player will not show correctly: a missing WEBVTT header, a timestamp that will not parse, cues out of order, a cue that ends before it starts, and two cues on screen at once outside a region with room to stack them. Every error names its line.
• Quality findings against a named captioning standard: line length, lines per cue, reading rate in characters per second, cues too short or too long, and gaps between cues short enough to flicker. Every finding names the standard and the limit, so the argument can be with the standard rather than with the tool.
• A machine-caption score: how likely the file is to be raw speech-recognition output nobody has corrected. A confidence with every signal listed and the evidence each one saw, never a bare yes or no.

Install

composer require bpmore/statamic-reel

Nothing else. The config file is optional:

php please vendor:publish --tag=statamic-reel-config

Reel adds a Captions tab to the edit screen of every video and audio asset, a Reel page under Tools, a reel:audit command, and two tags.

Caption tracks

Reel looks for a video's captions in two places, in this order:

1. The Caption track field on the asset, if somebody chose a file there.
2. A WebVTT file beside the video with the same name: talk.vtt or talk.en.vtt for talk.mp4, in the same folder. The code between the dots is the track's language. Several languages side by side give several tracks.

A chosen file that has since been deleted is reported as missing rather than quietly replaced by a sidecar, because somebody chose it.

Only WebVTT is read. An SRT file is not a caption track to a browser, and Reel will tell you its comma is not a full stop rather than guess.

The asset fields

On every video and audio asset, under Captions:

Set add_fields_to_assets to false in the config file to place the fields in your asset blueprints yourself, with the same handles.

The publish gate

Saving an entry that uses a video with no caption track, or a video whose track a player cannot read, or an audio file with no transcript, is refused, and the reasons appear under the field that holds the media. Three modes, overall and per collection:

'gate' => [
    'mode' => 'block',
    'collections' => [
        'blog' => 'warn',
        'drafts' => 'off',
    ],
],

block refuses. warn saves and writes the reasons to the log. off does nothing. An unrecognised mode is block: a typo must never be the reason a gate stopped gating.

The gate blocks only on what a file can prove. It never blocks on the machine-caption score, which is a likelihood, or on the audio description, which nothing can detect. Both are reported everywhere else.

The gate is standalone. It does not plug into Accessibility Gate's checks, which have no registration point; it uses the same refuse-with-reasons pattern so an author sees one kind of refusal.

The control panel

Tools, then Reel lists every media asset: its caption track, whether it has a transcript, the audio-description answer, and a verdict. Open an asset for the five criteria with the reason for each, and the caption editor.

The editor is a list of cues. Each has its identifier, start, end, settings and text, with buttons that move a timestamp by 100 milliseconds, and its findings listed beneath it. Findings update as you type. "Next finding" moves focus to the next cue with a problem. Deleting a cue takes two presses, the second labelled with the cue's number.

The whole editor works from the keyboard: every control is a real button or input in document order, nothing is hover-only, and nothing moves. Saving writes the WebVTT file back into the asset container through the same serializer the parser round-trips through, so notes, regions and styles you never touched are written back as they were. Saving needs the Edit caption files permission.

Tags

Reel ships no player. Use your own <video> or <audio> element.

<video controls src="{{ video:url }}">
    {{ reel:tracks :asset="video" }}
</video>
 
{{ reel:transcript :asset="video" }}

reel:tracks emits one <track> per caption track, with srclang and a label in the language's own name, the first marked default. Parameters: kind (captions by default; use subtitles for a track with no sound description), default="false" to mark none.

reel:transcript emits a <section> labelled by its own heading, with the transcript inside, or a link to the transcript entry. Parameters: heading (level, 2 by default), title (Transcript by default), class. It emits nothing at all when there is no transcript; it never fakes one.

Both accept an asset object, an asset id (assets::media/talk.mp4) or a path, and both work with no JavaScript.

The command

php please reel:audit
php please reel:audit --json | jq '.summary'
php please reel:audit --container=media --threshold=warn

Every media asset, its verdict, its criteria, and its caption file's errors, findings and machine score. --json is the whole of stdout and nothing else. Exit codes: 0 nothing at or above the threshold, 1 something is, 2 the command could not run. The threshold is fail by default and can be warn, manual or pass, from the config file or the option, so a team can start by failing the build on failures alone and tighten later.

Standards and thresholds

The quality findings are measured against one of three presets, netflix by default, each with a source in Bpmore\Reel\Vtt\QualityThresholds:

Where a standard gives words per minute rather than characters per second, or gives no figure at all, the config comments say what was derived and what was borrowed. Any figure can be overridden in the config file.

A gap between cues of 2 frames or less (83 ms) is a deliberate frame gap. A gap longer than that and shorter than half a second flickers and is reported.

Machine caption detection

Eight signals, each a ramp rather than a switch, weighted by how rarely a person produces it:

• No sentence punctuation and no capital letters: what recognisers used to leave out. All capitals is not a signal; no recogniser shouts.
• Uniform cue lengths: every cue the same duration and the same length, because a machine cut at a fixed window. Fast dialogue timed to a beat is uniform in one but not the other, and scores nothing.
• Word-level timing, lines repeated across cues and cues that open mid-sentence: what current recognisers leave in. YouTube's automatic track, checked September 2026, is punctuated and capitalised and gives itself away by timing every word and rolling each line through two cues.
• No non-speech annotations in ten minutes or more, and no speaker labels in what reads like a conversation: weak by nature and weighted so that a single voice at a lectern, professionally captioned, cannot reach "possible" on them alone.

The score is unlikely under 0.3, possible under 0.6, likely from 0.6. A likely score makes 1.2.2 Review, not a failure, and every signal is listed with what it saw so an editor can say "that was on purpose" and move on.

What it cannot see: a recogniser whose output has been lightly corrected, punctuated, capitalised and re-segmented at sentences looks like a person's work to these signals. Reel narrows the problem; it does not close it.

What Reel does not do

• Generate captions, recognise speech, or translate.
• Play media, host it, encode it, or burn subtitles in.
• Read SRT, TTML or SCC. WebVTT only, because that is what a browser plays.
• Tell a silent video from one with sound, or whether an audio description exists. Both are a person's answers, and Reel says so rather than guessing.
• Certify anything. A pass here is "nothing Reel can detect is wrong", and 1.2.4 is always a person's answer.

Support

Reel is sold on the Statamic Marketplace, one licence per production site, with local development and CI free under Statamic's own terms.

Bugs and questions go to GitHub Issues. A caption file the parser gets wrong is a bug: attach the file, or the shape of it, and it becomes a fixture. Issues are answered by the author; there is no service-level agreement.

ext-intl is optional. With it, each <track> is labelled in its own language ("Français"); without it, the label is the language code.

Developing

Pest, with the VTT engine tested without booting Statamic:

composer test

The addon's script is published into public/vendor/statamic-reel/js when it is installed. After editing resources/js/reel-editor.js on a site that installs the addon by path, run php please vendor:publish --tag=statamic-reel --force to see the change.