Subtitle generator

Burned in and separate are not the same thing

Subtitles come in two forms and people use the word for both, which causes a lot of confusion when choosing a tool. Burned in means the text is part of the picture, permanently. Everybody sees it, nobody can turn it off, and it works on every platform including the ones that strip caption tracks. That is what you want for a Short, because most of the people scrolling past have the sound off and no intention of tapping a settings menu.

A separate file, usually SRT or VTT, sits alongside the video. The platform draws it, the viewer can turn it off, search engines and platforms can read it, and you can edit it later without touching the video. That is what you want for a YouTube upload, a course, a client deliverable, or anything that needs to be searchable or translated later.

The reason this matters when choosing a tool is that plenty of them only do one. A tool that only burns in cannot give you a file for the long upload. A tool that only exports a file has not solved the problem for a Short. This one does both from the same transcript, which is also why they cannot disagree with each other.

Every transcript is wrong somewhere, so the fixing matters

Automatic transcription is very good and it is not perfect, and the places it fails are predictable: proper nouns, technical terms, brand names, people talking over each other, and anything sung. A tool that hands you a transcript and no way to correct it is a tool that guarantees a misspelled name on screen in front of an audience.

There are three ways to fix it here and they exist because they fail differently. You can retype any line directly, which is the fastest fix for one wrong word and keeps the timings. You can ask it to use the video’s own captions instead, which on music videos and most real uploads are written by a person and beat anything a machine produces from a compressed extract. Or you can paste the real words and have them timed to the audio, which is the answer for lyrics and for anything where you already have a script.

That last one is worth knowing about even if you never use it for music. If you read from a script, pasting it turns a transcription problem into an alignment problem, and alignment is a much easier problem. The words are then certainly right and only the timing is estimated.

How the subtitle file is built

Word timings are not subtitles. A file with one cue per word is technically valid and completely unreadable, so the words have to be grouped into lines the way a human editor would group them, and where the groups break decides how the file reads.

Four things close a cue here. A word ending in sentence punctuation, because a real sentence boundary is always the best place to break. A gap of about 280 milliseconds, which is roughly the pause you hear as the end of a thought. A duration cap, so no cue sits on screen long enough to stop being a cue. And a layout test, which closes a cue when the next word would no longer fit in two lines of 42 characters.

Those last numbers are the broadcast ones, used by the EBU and by Netflix among others, and they are not arbitrary. Two lines is what fits under a video without covering it, and 42 characters is roughly what a reader takes in per line at the speed speech arrives. Cues also get rebalanced so a stray word does not end up alone on a line of its own, which reads as a stutter.

Other languages, and the one most tools quietly skip

Captions can be translated into another language while the audio stays as it was, which is what you want for reaching an audience that does not share your language without redubbing anything. A translated line is timed as a whole line rather than word by word, so the highlight follows the sentence closely rather than exactly, and the page in the Studio says so rather than letting you find out.

Arabic and other right to left scripts render properly, with four Arabic faces built in, the letters joined and the line breaking handled in the correct direction. This sounds like a detail until you have seen the alternative, which is text arriving reversed or with every letter standing alone, and it is the single most common way a tool that claims a hundred languages turns out to mean a hundred Latin ones.

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