Subtitle Translator

The translation is done by Opus-MT, a family of compact neural machine translation models from the University of Helsinki, released under Apache-2.0 or CC-BY-4.0. Each model covers one language pair, which keeps it small: about 100–130 MB with 8-bit weights, downloaded once and cached. The tool offers more than 70 direct pairs, mostly to and from English plus several European pairs such as German–French or Spanish–Italian. When there's no direct model, it translates through English with two models, which works but can drift a little more.

Subtitle files are awkward to translate because sentences are often split across two or three cues. Translating each fragment on its own produces broken grammar, so by default the tool joins the cues of a sentence, translates the whole sentence, then shares the translation back across the original cues in proportion to their length, preferring breaks at punctuation. Timings never change. Two-speaker cues that start with dashes are translated line by line, cues wrapped in italics stay italic, and the original number of lines per cue is preserved with balanced line breaks.

ASS and SSA files from anime and fansub releases are handled in place: only the dialogue text changes, so styles, positions and leading override tags such as {\an8} survive, and you can export ASS again. SRT and VTT files can be exported in either format, and a bilingual option puts the translation above the original in italics, which is handy for language learners.

Everything runs in your browser on the CPU with WebAssembly, and the subtitles are never uploaded. Translation is fast for subtitle-sized text: a typical episode of 600–800 cues usually finishes in a few minutes on a laptop. Machine translation still makes mistakes with idioms, names, humour and gendered forms, so use the side-by-side editor to correct lines before publishing.