Can ChatGPT read SRT and VTT subtitle files?

Subtitle files are the least glamorous format on this site and one of the best behaved. Nothing to parse, no embedded images, no compression, no formatting layer to strip. Just text with some scaffolding around it.

What is inside an SRT file

Open one in any text editor and you get repeating blocks like this:

127
00:14:22,180 --> 00:14:25,640
The second quarter numbers came in
about four points under plan.

A cue number, a timing line, the words, then a blank line. A .vtt file is the same idea with small differences:

The test for any format is whether you can open it in a text editor and see readable words. Subtitle files pass it more plainly than almost anything else.

Why it uploads without trouble

OpenAI describes what it supports as "All common file extensions for text files, spreadsheets, presentations, and documents." A subtitle file is a text file, and both extensions go through in practice.

If one is ever refused, rename it to .txt. Nothing inside changes. The bytes are identical and only the extension differs, so this is not a conversion and costs you nothing.

The timestamps are most of the file

This is the part people miss.

Count the characters in one cue. A timing line like 00:14:22,180 --> 00:14:25,640 is 29 characters on its own, and the cue number and blank line add a few more. Subtitle text is deliberately short, often six to twelve words, because it has to fit on screen and be read at speed. So the scaffolding can easily match or exceed the dialogue it wraps.

That costs you twice. It burns length you might rather spend on content, and it feeds the model a steady stream of numbers that mean nothing to your question. Summaries of raw subtitle files commonly come back with timecodes sprinkled through them for no reason.

There is a subtler problem too. Subtitle lines break wherever they fit on screen, not where sentences end, so one sentence can be split across three cues. ChatGPT copes with this most of the time, but on dense technical material the broken lines make the text harder to follow than a clean transcript would be.

Stripping the timestamps

Two line types need to go: the one that is only a number, and the one containing the arrow.

1. Open the file in a text editor

   VS Code, Sublime, Notepad++, anything with find and replace across a whole file.

2. Turn on regular expressions

   Usually a small .* button beside the search box.

3. Delete the two structural line types

   Find ^\d+$\n and replace with nothing, then find ^.*-->.*$\n and replace with nothing. What is left is the dialogue.

4. Save it as .txt

   You now have a plain transcript, usually well under half the size of what you started with.

If you would rather not touch a regular expression, subtitle editing software can export plain text directly, and there are converters that do the same job in a browser.

When to keep the timestamps

They are not always noise. If you want to cite the moment something was said and jump back to it, if you are building chapter markers or a clip list, or if the order and spacing of statements matters to your question, the timings are the useful part of the file.

In that case leave them in and say so in your prompt. Tell ChatGPT it is looking at a subtitle file and ask for timecodes in the answer.

Subtitles as the video workaround

Video is not among the file types OpenAI lists for uploads. The four families named are text files, spreadsheets, presentations and documents, and video sits in none of them. Attaching an MP4 is not the route.

A subtitle file is. Most of the value in a lecture, an interview, a webinar or a deposition is in what was said, and that is exactly what the subtitle file holds. YouTube videos usually have a transcript you can open and copy. Downloaded videos often ship with an .srt beside them. Failing both, any transcription tool will produce one from the audio.

What you give up is everything visual. Slides, charts on screen, someone pointing at a diagram, an expression that changed the meaning of a sentence. If your question depends on what was shown rather than what was said, the subtitles will not answer it and nothing will warn you. Screenshot the frames that matter and upload those alongside the text.

Length, and where it actually bites

Speech runs at roughly ten thousand words an hour, so a single film or talk sits comfortably inside 2 million tokens. File size never matters here: subtitle files are tiny next to the 512 MB ceiling.

Volume is what bites. A full series, a term of recorded lectures, or a year of webinars pasted into one file can pass the length cap. Past it nothing is rejected. The file uploads, part of it is read, and no message tells you which part. Ask what happens in the final scene, or what the last speaker said. If it cannot answer, it did not get that far.

The short version

Upload the .srt or .vtt as it is for anything up to an hour or so. For a feature length film or a long course, delete the cue numbers and timing lines first, unless you want answers that cite timecodes, in which case leave them in and say so in the prompt. And when a video will not upload, which it will not, the subtitle file is how its content gets in.