Language and translation → Tools and technology

Tools and technology for Welsh

The tools that work, the gaps that remain, and the people building the next generation.

Four people exchanging colourful shapes through an abstract language machine

When word-for-word translation misses the mark

The Coleg Cymraeg Cenedlaethol report team tested Google Translate on common English idioms, accessing the service on 22 April 2025. The results clearly show the problem: translating word for word where a natural expression is needed:

English idiom	Google Translate translation
making a mountain out of a molehill	«gwneud mynydd allan o molehill»
on tenterhooks	«ar bachau tenter»
raining cats and dogs	«bwrw glaw cathod a chŵn»
to throw in the towel	«i daflu'r tywel i mewn»

The challenge, the report says, is not just translating individual expressions correctly, but reflecting what makes Welsh distinctive: the idiom, natural rhythm and cultural subtlety at the heart of a living language.

Where Wales leads: speech and transcription

Y Trawsgrifiwr, developed by Canolfan Bedwyr at Bangor University, is a resource for Welsh transcription. It transcribes Welsh lectures, interviews and multimedia content and — according to the report — offers advantages over commercial systems such as Otter.ai or Whisper, ‘which often fail to recognise Welsh because of a shortage of language data’ (Bangor NLP Group, 2022).

Through collaboration between the Welsh Government and Microsoft, meetings can now be transcribed in Teams in Welsh and English simultaneously. According to the Government’s summary of Welsh technology and AI for 2025–26, this is the first time this has been possible for any language pair.

Dedicated translation models: Techiaith and Dewi Bryn Jones

While major companies offer general translation services, Bangor University’s Language Technologies Unit (Techiaith) publishes fine-tuned translation models on Hugging Face, including:

nllb-200-1.3B-ft-eng-to-cym-legislation — fine-tuned on legislation, for the legal field;
mt-dspec-health-en-cy and mt-dspec-legislation-en-cy — models for the health and legislation domains;
m2m100_1.2B-ft-cy-to-en and nllb-200-1.3B-ft-cym-to-eng — for the reverse direction.

The account DewiBrynJones adds three further models:

bergamot-mt-cy-en — Welsh to English in Bergamot format;
m2m100_1.2B-ft-cy-to-en — Welsh to English;
nllb-200-1.3B-ft-eng-to-cym — English to Welsh.

Together, the two accounts contain eight translation models and ten translation datasets: translation memories from the Welsh Government, Cardiff University, BydTermCymru, legislation, health, YouTube subtitles and Bangor’s Transcription Bank. View all translation models and datasets →

Through the project Iriaith, funded by a Welsh Government technology grant to Bangor University, around 500,000 new Welsh–English sentence pairs were collected to develop machine translation and language models. The project also includes work on Y Trawsgrifiwr, text-to-speech technology and the Macsen digital assistant.

Try text-to-speech and speech recognition in your browser →

Established tools that remain useful

The report recommends using established Welsh resources alongside international AI tools: Cysill (spelling and grammar checker), Cysgeir (dictionaries), Termau Cymru (standard terminology database), and Macsen. For businesses and the public, the Welsh Government’s Helo Blod service maintains an up-to-date list of Welsh technology. A large language model can prepare a draft, but tools designed for Welsh should be used to check the language before publication.

A future risk: synthetic data

A warning from Techiaith: Welsh corpora collected from the web — such as OSCAR, HPLT and Common Crawl — already contain many machine translations and low-quality texts. The Coleg Cymraeg report identifies the next danger: a cycle in which models learn from content generated by other models, further degrading information quality. For a language with a small corpus, this is not a theoretical risk. That is why every authentic, human-written corpus — from the National Corpus of Contemporary Welsh to the National Library’s newspapers — is valuable to Welsh technology development.