Welsh translation and speech-to-text models cy

2 models into Welsh, 2 from Welsh, 12 speech-to-text. Best scored into Welsh: English to Welsh (chrF++ 28.7). Best scored from Welsh: Welsh to English (chrF++ 35.1).

Which open models work for Welsh?

Windy Translate lists 2 open translation models into Welsh and 2 from Welsh. All of them can be downloaded from Hugging Face and run on your own hardware. By licence: 2 Apache-2.0 and 2 CC-BY-4.0 (terms and attribution text: licensing). 2 of them have a screening score on FLORES-200. The best scored into Welsh is translate-en-cy (chrF++ 28.7, Limited). The best scored from Welsh is translate-cy-en (chrF++ 35.1, Usable).

With English. English to Welsh: translate-en-cy (chrF++ 28.7, Limited), the best of 2 direct models. Welsh to English: translate-cy-en (chrF++ 35.1, Usable), the best of 2 direct models. To translate between Welsh and a language with no direct model, you can go through English with these two.

By band, the scored Welsh models are 1 Usable and 1 Limited (Excellent is chrF++ 60 or more, Good 45 to 60, Usable 30 to 45, Limited 15 to 30). A score is a 48-sentence screening of news and encyclopedia text, not a measure of accuracy, and a difference of a point or two is noise. How we measured.

Speech-to-text. Whisper's tokenizer lists Welsh as one of its languages, so all 12 multilingual Whisper-based models here can transcribe it: see the speech models. Accuracy differs a lot between languages and model sizes, and this site has not measured speech accuracy.

Each model page has a short snippet to download and run the model yourself, its licence and attribution text, and its score where there is one. Pick the best scored model for your direction, then try it on a few sentences of your own text: a screening score is a starting point, not a guarantee.

Into Welsh

From Welsh

Scores: chrF++ on FLORES-200 dev, 48 sentences, measured 2026-08-04. Method.