How to export an SRT file from a video.
Three routes, and the two rules that break hand-made files.

The word for this is a sidecar: a separate file that travels next to the video instead of inside it. That separation is the point. The picture is never touched, so the same master can carry an English file, a French file and a corrected version of both without a single re-export. A player reads the file, matches the timecodes to its own clock, and draws the text on top.

It also explains the request you are probably answering. A broadcaster, a social team or a web developer asks for the SRT because they want to control what happens to the text: translate it, restyle it, index it, or turn it off. Hand them a video with the words burned into the frames and they can do none of that.

One consequence catches people out. Because the file is only text and timecodes, there is nothing to extract from a video that has no caption track. If the words were burned in at export they are pixels now, and the only honest route back to a file is to transcribe the audio again.

Nearly every broken subtitle file fails on one of two things, and both are invisible until a player refuses the file with no explanation.

The first is the decimal separator. SRT writes a comma before the three digits of milliseconds. WebVTT, its web-native cousin, writes a full stop: the W3C specification is explicit that a cue timestamp takes a full stop character followed by three ASCII digits representing thousandths of a second. The two formats are close enough that people copy a timestamp from one into the other, and the file then looks correct to a human and wrong to a parser. If a file will not load, look at that character before you look at anything else.

The second is the blank line. Every cue ends with an empty line, and the parser uses it to know where one cue stops and the next begins. Delete a cue by selecting its text and the blank line usually goes with it, which welds two cues together into a block with two timecode lines. Some players skip the block, some skip the rest of the file. Neither tells you why.

A third habit is worth adopting even though nothing enforces it. Save as UTF-8, without a byte order mark, and keep the line endings consistent. Accented characters are where a file that works on your machine turns into question marks on somebody else’s.

Route one is first for a reason that has nothing to do with tooling: the timings came from the same timeline as the cut, so they cannot drift. Route two produces timings from the audio, which is usually close and occasionally a beat late. Route three produces whatever you were paying attention to at the time.

Whichever route you take, generate or export after picture lock. A two-second trim in the middle of the film moves every cue after it, and there is no repair for that which is faster than starting the track again.

Line length and reading rate are the same problem seen from two angles. A cue that runs to a third line is almost always a cue that should have been two cues, and splitting it fixes the length and the reading rate together. Break at a natural phrase boundary, not at the point where the box happened to fill.

Speaker changes are the other thing an SRT can express and most machine output does not. When two people speak inside one cue, the convention is a hyphen at the start of each line so the reader can see the exchange. It matters most in the places automatic transcription is weakest: interviews, overlapping dialogue, a question and its answer landing inside two seconds of each other. If the file is going to a broadcaster, this is a compliance item, not a preference.

Then there is timing feel. A cue that arrives slightly early is invisible; a cue that arrives slightly late makes the viewer feel they are being told what they just heard. When a cue sits off, pull it earlier rather than pushing it later. And end a cue on a hard cut rather than letting it run into the next scene, where the eye attaches the words to the wrong image.

That is also the reason to keep SRT as your default deliverable. It expresses less than the alternatives, and expressing less is what makes it portable: there is nothing in the file for a receiving system to misinterpret. When someone asks for a format you do not produce, sending the SRT and saying plainly that it is the file to convert from is a complete answer, because almost every subtitle tool takes SRT as an input.

The exception is the web page. If the film is going into an embed on a site, whoever builds the page wants the VTT, and sending it unasked saves a round trip.

Two details worth knowing before you send the file on. WebVTT can carry inline tags that SRT has no equivalent for, so they are stripped during the conversion rather than passed through as visible junk in the text. And the file arrives named after the film, so a delivery folder does not fill up with four files called captions.srt.

Generation is charged against a monthly minute allowance that grows with the plan, so the thing to check before starting a long film is the runtime against the allowance. Adding subtitles to a client video walks the whole loop in order, and captions for client work covers what to hand over at the end of a job.