Voice
Try technology that speaks Welsh and recognises spoken Welsh, and see how we plan to measure its quality.
These speech models run on your device. After downloading the models, processing takes place in your browser and the demo sends no text or recordings to AIsteddfod. The first download may take a little time, but your browser will retain the speech model for reuse.
Text-to-speech
Type something in Welsh. The CC0 model Techiaith’s cy_GB-bu_tts generates audio on your device through Piper/VITS, ONNX and WebAssembly; you can choose three of the model’s seven voices. eSpeak WASM, which converts text into speech sounds, is licensed under GPL-3.0-or-later. Not all components of the speech system therefore have permissive licences.
Welsh CC0 speech model not loaded (around 77 MB).
Speech-to-text
Speak in Welsh. The BangorAI Zipformer research model transcribes as you speak through sherpa-onnx WASM, without sending audio to a cloud service. This is an interim model while we develop the new browser model.
BangorAI Welsh Zipformer model not loaded (around 86 MB).
Demo details: the interim Zipformer model achieved a word error rate (WER) of 24.57% on the development set and 25.70% on the held-out test set. The streaming version achieved 43.87% WER on ARFOR test-clean v0.1 before changing how audio segments were finalised. On the ARFOR development set only, WER fell to 27.50% across all 3,441 clips when finalising audio segments every three seconds (RTF 0.090). The model was not retested on the test set after the change. In a separate provisional benchmark, Techiaith’s Whisper model was more accurate; these scores cannot be compared directly with Techiaith’s published protocol. Read the model card →
What needs measuring?
Transcribing words or generating audio is not enough in itself. Voice technology must work with Welsh as it is spoken: at different speeds, in different accents, and alongside English when people switch languages mid-sentence.
Speech recognition
We will measure word error rate and check how the system handles personal and place names, punctuation, language switching and different accents.
Text-to-speech
Listeners will assess clarity, pronunciation, stress, rhythm and naturalness — including numbers, dates, abbreviations and mutated words.
Practical use
Tests will include clear and noisy audio, short sentences and longer passages, and tasks such as recording a meeting or reading instructions.
The planned protocol
- The same material for every system. A fixed set of recordings and scripts will be used, with clear rights and consent.
- A variety of voices. The set will include different accents, ages, speaking speeds and bilingual examples.
- Automatic measures and human judgement. Word error rate is useful for transcription; for synthetic speech, intelligibility and naturalness are equally important.
- Publishing the details. The system version, settings, test material and scoring method will accompany every result.
Available Welsh tools
Several Welsh voice resources are already available. Wales’s National Language Technologies Portal offers resources for text-to-speech and speech recognition, and the Trawsgrifiwr converts spoken Welsh into text. The Welsh Government also records ongoing work on Welsh speech technologies and AI.
35 speech models and 25 audio datasets in the catalogue. The list includes public items from the Hugging Face accounts Techiaith and DewiBrynJones, including Whisper, Kaldi and Wav2Vec2 series, text-to-speech models, evaluation datasets and live demos. Open the voice and speech catalogue →
See also the language tools overview and the resource directory.