Task	Description
Welsh transcription	Welsh audio → Welsh text
English transcription	English audio (UK/Irish accents) → English text
Welsh→English translation (experimental)	Welsh audio → English text — limited quality, not recommended for production (see below)

Evaluation Results

WER / CER (lower is better) on the held-out benchmark. Average across the 4 transcription test sets: WER 20.29, CER 7.33.

Welsh Transcription

Test Set	WER	CER
cymen-arfor/lleisiau-arfor (spontaneous)	29.20	11.59
techiaith/banc-trawsgrifiadau-bangor (mixed)	26.71	9.91
techiaith/commonvoice-23-0-cy (read)	15.10	4.35

English Transcription

Test Set	WER	CER
techiaith/commonvoice-23-0-en/GB-IE (read, UK/Irish)	10.16	3.47

Welsh→English Translation

The model can translate Welsh speech directly into English text (task="translate"). This capability is experimental and has not been formally evaluated; its translation quality is limited and it is not recommended for production use.

For higher-quality English, transcribe the Welsh audio with this model and translate the resulting Welsh text with a dedicated Welsh→English machine translation model.

Training Data

Dataset	Language	Split	Revision
techiaith/banc-trawsgrifiadau-bangor	Welsh	`train`	`26.07`
techiaith/corpws-clllc-wlga	Welsh	`clips`	`5d00294c31`
techiaith/commonvoice_23_0_cy	Welsh	`train+dev[50%:]`	`main` †
cymen-arfor/lleisiau-arfor	Welsh	`train`	`main`
techiaith/commonvoice_vad_cy	Welsh	`train`	`main`
techiaith/commonvoice_23_0_en__GB_IE	English	`train[:10%]`	`main` †
techiaith/commonvoice_23_0_cy_en	Welsh	`train[:10%]`	`main` †

† Common Voice data is not redistributed here. techiaith Common Voice 23.0 data is held in private repositories, in line with the Mozilla Foundation's terms for obtaining and sharing Common Voice data. They are listed above for provenance and reproducibility, not as downloads.

The Welsh Common Voice corpus is freely available from Mozilla themselves: Mozilla Common Voice — Welsh datasets. Releases are distributed through the Mozilla Data Collective, and the datacollective package offers a programmatic download. With that in hand the rest of this pipeline is reproducible — every other dataset in the table is openly available on the Hugging Face Hub. Our thanks to Mozilla and to the Welsh speakers who contributed their voices.

Training Configuration

Parameter	Value
Base model	openai/whisper-large-v2
Learning rate	1e-05
LR scheduler	cosine
Warmup steps	500
Max steps	15000
Weight decay	0.01
Batch size	16 × 2 accumulation × 2 GPUs = 64 effective
FP16	True
SpecAugment	True
Curation applied	True
Early stopping patience	10
Selection metric	eval_btb_wer

Usage

python

from transformers import pipeline

pipe = pipeline(
    "automatic-speech-recognition",
    model="techiaith/whisper-large-ft-cy-en",
)

# Welsh transcription
result = pipe("welsh_audio.wav", generate_kwargs={"language": "cy", "task": "transcribe"})

# English transcription
result = pipe("english_audio.wav", generate_kwargs={"language": "en", "task": "transcribe"})

# Welsh to English translation
result = pipe("welsh_audio.wav", generate_kwargs={"language": "cy", "task": "translate"})

CTranslate2 Version

A CTranslate2 (int8 quantised) version is available at techiaith/whisper-large-ft-cy-en-ct2 for faster inference.

Acknowledgements

Developed by Uned Technolegau Iaith, Prifysgol Bangor / Language Technologies Unit, Bangor University.

Funded by the Welsh Government.