Can AI Speak Welsh? A Learner's Guide to Welsh Language Technology

Ask an AI tool to translate a sentence into Cymraeg and it may answer in a second. Ask it to explain a mutation, recognise fast regional speech and keep the same natural register, and the confidence of the answer can hide much weaker ground.

Welsh-language technology is real and useful. Researchers have built speech corpora, voice assistants, text tools and evaluation methods for years. The best learner question is not whether AI speaks Welsh at all, but which task it can support and where a human or trusted dictionary should take over.

Can AI speak Welsh? What current tools already do

Welsh language technology includes spellchecking, machine translation, speech recognition, speech synthesis, predictive text and conversational systems. Bangor University researchers and partners have developed resources that help both public tools and further research.

A voice assistant called Macsen demonstrated how Welsh speech, language processing and services could be joined into a useful interaction. The significance is not that one assistant solved every problem. It showed what becomes possible when Welsh data and Welsh-speaking users are included deliberately.

Why Welsh is described as a less-resourced language

Large AI systems learn patterns from data. English has enormous volumes of digitised text, labelled recordings, subtitles and commercial investment. Welsh has far fewer resources of the same scale, balance and technical preparation.

Less-resourced does not mean less expressive or less suitable for technology. It describes the data and infrastructure available to developers. A smaller dataset makes careful collection, consent, dialect coverage and public access more important.

Why speech is harder than typed Welsh

Clean, read speech is easier for a recognition system than overlapping conversation in a noisy kitchen. Natural speakers shorten words, switch pace, use local vocabulary and carry regional sound patterns. A microphone and room add another layer.

The crowdsourced Paldaruo corpus was created to expand Welsh speech data. Research comparing speech conditions has shown why a promising score on prepared audio should not be treated as a guarantee for spontaneous speech.

Dialect is not an error to be corrected away

North and south Wales varieties can differ in vocabulary, pronunciation and common structures. Individual speakers also move between formal and informal registers. A system trained unevenly may label a perfectly natural form as unusual or replace it with a more frequent alternative.

If a tool changes rŵan to nawr, for example, the useful question is not simply which is correct. Both can mean now. Ask which variety, speaker and situation the form belongs to.

A safe four-step workflow for learners

1. Attempt: write or say the Welsh you can produce yourself.
2. Compare: ask the tool for one alternative, not ten overwhelming rewrites.
3. Verify: check the key word in Geiriadur Prifysgol Cymru, TermCymru or your course materials.
4. Retrieve: close the answer and produce the sentence again from memory.

This keeps AI in the role of practice partner. Retrieval, not copying, is what turns the encounter into learning.

Prompts that expose uncertainty

Ask a tool to label north, south, formal and informal alternatives. Request a short explanation of any mutation. Then ask it to state which words it is least sure about. These prompts cannot force honesty, but they make unsupported certainty easier to notice.

For pronunciation, request a real audio source from a Welsh broadcaster or dictionary instead of relying only on invented English spelling. The phrase Dw i ddim yn siŵr means I am not sure. A good learning system should leave room for that sentence.

Think before uploading a voice recording

A child's voice, a private lesson and a workplace conversation can contain personal information. Read the service's privacy terms before uploading audio. Check whether recordings are retained, used to improve models or shared with another provider.

For low-risk practice, record a sentence created for the exercise rather than a real conversation. Delete the file when it has served its purpose. Language learning does not require trading away someone else's privacy.

What better Welsh AI would look like

Better systems need more than a larger model. They need consented Welsh data, transparent evaluation, coverage across dialects and ages, accessible tools and sustained work with Welsh-speaking communities.

Success should include the ability to preserve a speaker's chosen variety, admit uncertainty and support people who are learning. The goal is not for technology to replace siaradwyr - speakers. It is for speakers to shape the technology built around their language.

Keep technology tied to real practice with our everyday Welsh routine, then see how one word changes across contexts in our guide to hwyl.