Translate ASS and SSA subtitles without breaking the typesetting

Why ASS files trip up general AI assistants

An ASS script is not a list of sentences. Each line of dialogue is an event such as "Dialogue: 0,0:03:12.40,0:03:15.02,Default,,0,0,0,,{\an8}Text", with ten comma-separated fields, and the text itself often contains braces full of rendering instructions: a position, a fade, a color, a font change halfway through a word. Above the events sit the script header, the style definitions and sometimes whole fonts embedded as text.

To translate that in a chat, an assistant has to write the entire file back out. On an episode it tends to stop before the end, and along the way it may translate words inside the braces, drop a tag it does not recognize, or shift a comma so a field moves. Any of these changes how the line renders, and some break it completely.

This tool never asks a model to rewrite the file. It reads the script, takes only the Text of each Dialogue line, and keeps the tags out of the model: blocks at the start and end of a line stay aside, and blocks in the middle become numbered placeholders that are put back after translation. The text goes out in blocks of up to 60 lines, each line must come back under its own number, a missing line is asked for once more, and if it is still missing the original stays in place. Everything else in the file is copied as it was.

What is translated and what is left alone

• Translated: the text of every Dialogue line that has words, including signs and on-screen text typeset with \pos, \move or \frz
• Left alone: [Script Info], styles, [Fonts] and [Graphics] sections, Comment lines, actor names and the Effect field
• Left alone: drawing lines ({\p1} vector shapes) and lines with no words
• Left alone by default: karaoke lines with {\k}, {\kf} or {\ko} syllable timing, usually romaji song lyrics. Choose "Translate them" and the \k tags are removed, the line is translated, and it keeps its start and end time
• Bilingual option: the original line above the translation, in the same event, so both show with the same style and position

What people translate

• Fansub releases of anime, into a language the group does not cover
• A full cour of 12 or 24 episodes for friends or family who read another language
• Donghua and drama scripts typeset in Aegisub
• Their own subtitles for a video, shared in more languages with the styling kept
• Bilingual scripts for learning Japanese, Chinese or Korean while watching

Limits and getting a good result

Typesetting is kept, but it was made for the original words. A sign fitted tightly into a box or masked with \clip can overflow when the translation is longer, and copies of the same sign stacked on several layers are translated line by line, so check signs in the first part and fix any you need in the editor. The fonts named in your styles stay too; if they have no letters for your target language, such as Arabic or Chinese, the player may fall back to another font or show empty boxes, so change the Fontname in the style section if that happens.

Karaoke is recognized by its {\k} tags. Effects already generated by a karaoke template, one syllable per line, are not recognized as karaoke; translate the script before applying the template, or remove those lines first. Image-based subtitles (.sup, .sub with .idx) and video files are not supported, and the timings are not changed: if the script is out of sync with your copy of the video, the translation will be too.

It is machine translation, well suited to watching and sharing. For a release under your group's name, have a translator or editor read it.