SRT and VTT subtitles: one more file, not one more invoice

Say "automatic subtitling" and the picture that comes to mind is an AI that listens back to a recording and guesses what was said. That is what PlayPlay sells, and what most tools that stamp the word "AI" on this feature charge for. Here, the subtitle is not guessed: it is already there. It is a by-product of what the player does anyway, and that difference is not marketing trivia. It changes what the feature costs and what you can rely on.

Why they cost nothing to make

When an article is narrated, the engine does not just return an audio file. It also keeps, for every single word, the instant it starts and the instant it ends. That word-level timing is what drives the synchronised highlighting of the text during playback, the feature described in the piece on guided reading and in the one on word-level timestamps. And a subtitle, at bottom, is exactly that: a chunk of text, a start time, an end time.

So building an SRT or a VTT needs no new processing. No extra AI call, no model listening back to the audio, no additional storage. The file is assembled on the fly from data already captured on the day the audio was generated. That is why the feature is free where others sell it as a separate module: we have nothing to re-bill, because we have nothing to re-compute.

The one condition, stated plainly

There is a limit, and it deserves to be put in writing rather than discovered in use. Word-level timing only exists for audio generated from 13 July onward. An article narrated before that date has no synchronisation track: it has a sound file, not the list of words with their timings. For that article, the subtitle export does not return an empty file that would look like a bug; it returns a clear message saying what is missing and how to get it. Regenerating the article's audio recreates the timing, and the subtitles follow.

This is an acknowledged debt, not a hidden fault. We would rather tell you "this audio is too old, regenerate it" than hand you an approximate subtitle file rebuilt by a separate transcription, which would carry its own margin of error and would sometimes contradict what the voice actually says.

Two formats, two uses

The export offers both SRT and VTT, and the choice is not cosmetic. SRT is the universal subtitle format: a video platform, an editing suite, a media player all understand it with no setup. VTT (WebVTT) is the web format: it is what a track tag expects inside an HTML video, and it carries the language information, which matters when a page mixes several languages. The only visible difference between the two is a matter of convention, a comma or a full stop before the milliseconds in the timecode, but each ecosystem is strict about its own.

What separates a readable subtitle from a merely valid one

A subtitle file can be perfectly valid and still be unreadable on screen. That is the trap of tools that cut on the word or the round second. We apply the conventions of professional subtitling, the ones used by the BBC, Netflix or broadcast regulators, because accessibility is what these files serve, and a subtitle nobody can follow helps nobody.

In practice, that means a few simple, non-negotiable rules:

• A line never exceeds forty-two characters. Beyond that, the eye has to travel horizontally and drops out.
• A subtitle fits on two lines at most. A third hides the image and runs past the reading span.
• A subtitle stays on screen for at least one second, or it flickers past unread, and for at most seven, or the eye rereads it and loses interest.
• The cut falls on a natural boundary of the text: the end of a sentence, a breath marked by a comma, a real silence between two words. Never in the middle of an idea.

These rules are not there to look tidy. They are what separates a file you can drop onto a video from a file you would have to fix by hand.

What it is actually for

The first use is accessibility, in the strict sense. A deaf or hard-of-hearing person gets no benefit from the audio version of an article; the subtitle makes that content legible at the same time as others listen to it. The second use is video: a sound clip laid over a still image, a cutdown for social feeds, a short montage are most often watched with the sound off, and a correct subtitle is what makes people watch them to the end.

The export can be pulled from the interface for a given article, and through the API for anyone automating their publishing chain, in the logic set out on the French text to speech page: the same data serves the voice, the highlighting and the subtitle, because it was captured once, cleanly, at the right moment. The rest of the player's behaviour is detailed on the dedicated subtitles page.

Subtitling is not one more AI module. It is what timing data becomes when you have bothered to keep it. The quality rests on the accuracy of that data, not on the power of a model guessing after the fact: the subtitle says exactly what the voice pronounces, at the right moment, because it comes from the same source.

Read next

• The œ ligature: the little sign that gives a French speech engine away
• A priori, in fine, etc.: how a French voice should read Latin
• Adding an audio version to your Craft CMS articles