SRT generator

What is actually inside an SRT file

SRT is about as simple as a file format gets, which is why it has outlived every proprietary alternative that tried to replace it. Each cue is four things: a number counting up from one, a line holding a start and end timecode separated by an arrow, one or two lines of text, and a blank line. That is the entire specification in practice.

The timecodes are written as hours, minutes, seconds and milliseconds, with a comma before the milliseconds rather than a full stop. That comma is the single most common reason a file gets rejected, because VTT uses a full stop in the same position and the two formats otherwise look identical enough to confuse anybody writing one by hand.

Because it is plain text, you can open an SRT in any editor and fix a typo without regenerating anything, which is genuinely useful and also the reason a badly generated file stays badly generated until somebody notices. Getting the cue grouping right at the point of generation saves the work that would otherwise happen at two in the morning before a deadline.

Why one cue per word is a broken file

Speech recognition gives you a start and end time for every word. Writing those straight out produces a technically valid SRT that no human can read, because the text flickers a word at a time and there is nothing to fix your eye on. Grouping is the entire craft of subtitle generation and it is what separates a usable file from an export.

Four signals close a cue here. Sentence punctuation, because a real sentence boundary is always the best available place to break. A gap of about 280 milliseconds, which is roughly the silence a listener hears as the end of a thought. A duration cap, because a cue that sits too long has stopped functioning as a cue. And a layout test that closes the cue when the next word would no longer fit inside two lines of 42 characters.

That last one is deliberately a layout test rather than a character count, and the difference is not academic. Capping a cue at 84 characters looks equivalent to two lines of 42 and is not: 84 characters whose word boundaries fall badly have no split where both halves fit, and the result is a line that overflows. Testing the actual layout means a cue that cannot be laid out is never built in the first place.

There is one more pass afterwards. A cue left holding a single stray word reads as a stutter, so words walk backwards across the boundary one at a time until it is respectable, and only ever across a boundary this process created. A cue that ended because the speaker finished a sentence is never touched, because that boundary came from a person and is more correct than anything a rule could compute.

The two things that break subtitle files

The first is encoding. Files here are UTF-8 with no byte order mark. Some very old players wanted a mark at the front and some parsers read it as part of the first cue number, and since every player anybody actually uses now reads UTF-8 without one, adding it solves a problem that expired a decade ago while creating a live one.

The second is a mismatch between the file and the video it belongs to. If you trim the video after generating the file, every timecode is now wrong by the amount you trimmed off the front. Generate the file after the cut, not before, and if you must do it the other way round, expect to shift the whole file.

Where an SRT actually goes

On YouTube it goes in the subtitles section of a video you have already uploaded, and it is worth doing on long uploads specifically because the platform can read it, which burned in captions being pixels cannot offer. On Premiere, Resolve and Final Cut it imports as a caption track you can restyle. On a website it works but VTT is the format the browser actually wants, which is why both come out of the same transcript here.

For a Short, a separate file is usually the wrong tool. Most people watching have the sound off and are not going to tap a settings menu, and some platforms strip caption tracks entirely, so captions burned into the picture are the ones that get read. Doing both is common: burn them in for the Short, keep the file for the long upload.

How to do it

Specifications

The numbers rather than the adjectives, including the limits. If one of these is a dealbreaker it is better learned here than after paying.

What people ask about this

Keep going