18 September 2026
From meetings to evidence: a clearer map for Welsh AI
There was no new Welsh model this week, but practical use in the Senedd and work on lower-resource languages show how to move from data to a trustworthy test.
A Welsh meeting in minutes
In evidence to a Senedd committee on 14 September, Dafydd Trystan Davies said it was possible to hold a meeting in Welsh on Teams, transcribe it, and produce a bilingual note ‘quite well’ within minutes.
That is a practical sign of value. This is not just a laboratory experiment, but a familiar public administration task: listening to, recording and summarising a Welsh meeting.
But the description is not a benchmark. The record does not say which Teams feature or version was used, how much of the conversation was in Welsh, how much correction was needed, or whether names, dates, amounts and decisions were preserved accurately. That is the important gap: we have an indication that the method is useful, but not the evidence to say when it can be trusted.
A small, careful test would be far more valuable than a broad claim. The audio, raw transcript, generated note and human-corrected version could be retained. We could then measure word errors, speaker recognition, switching between Welsh and English, correction time, and mistakes that change meaning.
The hidden layer: the tokeniser
The strongest technical lesson came from Nemotron-SEA-LION v4.8, a new family of four 30B and 120B models for Southeast Asian languages. The weights are downloadable, and the models have improved on SEA-HELM, a benchmark developed with native-speaker input.
The 30B model's score rose from 46.89 to 51.57 compared with its corresponding Nemotron base. At 120B, the score rose from 49.30 to 63.44. The gains were particularly large for Burmese and Tamil.
But the limitation is the most useful part for Wales. The team retained Nemotron's original tokeniser. For Khmer, Lao and Tamil, it uses around five times as many tokens as SEA-LION's previous tokeniser to represent the same text.
A tokeniser breaks text into pieces the model can process. If Welsh is broken into too many pieces, the same paragraph can consume more of the context window, cost more and provide less information per training step. Adding lots of data can help, but cannot remove every disadvantage in the underlying representation.
Every Welsh model card should therefore answer a simple question: how many tokens does this text need compared with English? That is not an appendix detail. It is a measure of cost, speed and capacity.
There is no need to rebuild everything
On 15 September, a Whisper-small model for Cebuanowas updated. Its card records a word error rate of 5.10% on the validation set, compared with 11.14% in the earlier run. The character error rate fell from 4.51% to 2.59%.
The 241.7-million-parameter weights are available under Apache-2.0. But this is the team's own result on in-domain data, not an independent test on all types of Cebuano speech. The lesson is that continuing to train a small specialist model can pay off—not that Cebuano has been ‘solved’.
In Sequential Adapter Stacking, the researchers placed a related-language adapter beneath a new target-language adapter. On Asturian, Assamese and Xhosa—three languages Whisper did not support—the method was 5–8% better than full fine-tuning in relative word error reduction.
With just one hour of target data, error fell by 12% for Assamese and 26% for Xhosa compared with full fine-tuning. No Welsh model has been released from this work, and we should not assume in advance which language would be the best source for Welsh. But it offers a clear experiment: compare full fine-tuning, a conventional adapter and transfer from several possible languages under the same data budget.
From audio to a correct answer
VākQA shows how to test more than transcription. The new set contains 2,001 Telugu question–answer pairs, 2.53 hours of audio, bilingual transcripts and six subject areas. The data is publicly available, and people have checked the questions and answers.
When Gemini received correct Telugu text, its average score was 3.63 out of 5. Answering directly from audio, it scored 3.28. Compared with the correct Telugu text, 21.2% of questions worsened, 13.1% improved, and 65.7% were unchanged.
This separates three questions that are often confused:
- did the system hear the words correctly?
- did a translation preserve the meaning and local context?
- could the model reason its way to the correct answer?
The same distinction is essential for a Welsh meeting note. A transcript can contain several small errors but preserve the decision; or read smoothly while an important name, amount or commitment has changed.
TeochewBench adds a cultural warning. On 300 human-checked utterances, leading models' average score fell from 69.25 on the least specific items to 27.52 on those most specific to Teochew culture. Superficial text overlap can make a model look better than it is, so the researchers also used a copying control.
For Welsh, a test should have separate layers for literal sentences, idioms, local names, cultural references and local knowledge. It should also include a simple baseline to show whether the model understands or copies.
Context on the watchlist
On 11 September—the day before this report's window—the Welsh Government announced a partnership with Cosine. The intention is to put Welsh data, resources and expertise into Lumen Sovereign's training pipeline.
It is a promising development, but there was no new update this week. The announcement does not yet name the corpora, data volumes, licences, Welsh test or release timetable. We should welcome the investment and ask for the details at the same time. The project will remain on our watchlist until there is an inspectable artifact or result.
The next step for Wales
The pieces now fit together. The Senedd evidence provides the real task: a Welsh meeting and a bilingual note. SEA-LION shows that the tokeniser affects cost and context. VākQA separates listening, translation and reasoning. Cebuano and stacked adapters show that small, targeted experiments can improve an existing system.
The most useful contribution now would be a repeatable public test:
- meeting samples with consent, including code-switching and a range of accents;
- separate scores for speakers, language, words, summaries and critical facts;
- a comparison of Welsh token costs and lengths;
- fixed model versions and published normalisation rules;
- and a human check of the final meaning.
This was not an empty week for Welsh AI. It was a week without a new Welsh launch, but with a clearer map of what needs to be built and measured next.