Skip to content
Cutshed
Menu

Subtitles & text · 23

Subtitle translator

Each cue is translated on its own, so the timestamps you already timed stay on the same lines. The first pair downloads an OPUS-MT model, on the order of 80 MB, and caches it in the browser.

Each pair is an OPUS-MT model from Helsinki-NLP, CC BY 4.0, fetched from Hugging Face the first time you pick it (about 80 MB, quantized) and cached in this browser. NLLB is not used: the distilled weights are CC BY-NC. Portuguese is not in the list because there is no small dedicated ONNX pair for it on that hub. The subtitle file itself is not uploaded.

How to use it

English, Spanish, French, German, and Italian are the pairs, in both directions. Portuguese is not in the list: the public Xenova opus-mt English–Portuguese repos are not available, and the Romance many-to-one model needs language tokens this page does not send.

NLLB is not used. The distilled NLLB weights are CC BY-NC, which is not a license for a tool people run on commercial videos. OPUS-MT from Helsinki-NLP is CC BY 4.0, which allows commercial use with attribution.

Four hundred cues is the cap. A feature-length subtitle file will time out a laptop. Translate a reel, check the lines, and expect names and jokes to need a human pass. The model is sentence-level machine translation, not a dialogue editor.

Tips

  • Keep cues short before you translate. A cue that is three sentences becomes one long line and will not fit a caption box.
  • Do not translate twice. English to Spanish to English is not a cleanup pass.
  • Read the preview out loud. Proper nouns are the usual failure.

FAQ

Which model runs?

A Helsinki-NLP OPUS-MT pair via Transformers.js, 8-bit, on WASM. English–Spanish, English–French, English–German, English–Italian, and the reverse directions.

Why is Portuguese missing?

There is no usable public Xenova opus-mt English–Portuguese model for this runtime. Shipping a non-commercial NLLB checkpoint instead would break the license rule.

Do the timestamps change?

No. Only the cue text is replaced. Use the timing shifter if the whole file is early or late.

License

OPUS-MT models published by Helsinki-NLP are Creative Commons Attribution 4.0. Attribute the Language Technology Research Group at the University of Helsinki. Transformers.js and the ONNX runtime it uses are Apache-2.0 and MIT. NLLB is not included because the distilled weights are CC BY-NC.