BuzzASR/welsh

1---2language: cy3license: mit4library_name: transformers5pipeline_tag: automatic-speech-recognition6base_model: openai/whisper-large-v37tags: [automatic-speech-recognition, whisper, welsh, buzzasr]8datasets: [google/fleurs]9metrics: [cer, wer]10---11 12# BuzzASR — Welsh13 14A monolingual automatic speech recognition model for **Welsh**, fine-tuned from15[openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3). Part of **BuzzASR**,16a suite of 102 language-specialized ASR models17([paper: arXiv:2609.09554](https://arxiv.org/abs/2609.09554), Findings of EMNLP 2026).18 19This model uses **full fine-tuning (native per-language tokenizer replacement + text multitask fine-tuning)**.20 21> 🏆 **State-of-the-art (open-source).** On the combined FLEURS + Common Voice test set, this model22> achieves the lowest CER of every open system we compare against: Whisper-large-v3, Omnilingual 1B/7B, MMS, Qwen3-ASR, and Cohere Transcribe.23 24## Results (normalized CER / WER, %)25 26| Test set | CER | WER | Whisper-large-v3 (zero-shot) CER |27|---|---|---|---|28| FLEURS | 7.31 | 17.45 | 11.36 |29| Common Voice 25 | 2.25 | 5.98 | 13.73 |30| Combined | 5.24 | 12.57 | 12.88 |31 32~2.5x CER reduction over Whisper zero-shot on the combined test set.33 34## Usage35 36```python37import torch, torchaudio38from transformers import WhisperForConditionalGeneration, WhisperProcessor39 40model = WhisperForConditionalGeneration.from_pretrained("BuzzASR/welsh", torch_dtype=torch.float16).to("cuda").eval()41proc  = WhisperProcessor.from_pretrained("BuzzASR/welsh")42 43wav, sr = torchaudio.load("audio.wav")           # 16 kHz mono44feats = proc(wav[0], sampling_rate=16000, return_tensors="pt").input_features.to("cuda").half()45ids = model.generate(feats, num_beams=1, no_repeat_ngram_size=3, repetition_penalty=1.2)46print(proc.batch_decode(ids, skip_special_tokens=True)[0])47```48The language/task prompt is baked into the generation config, so no `language=` argument is needed.49 50## Training data51[FLEURS](https://huggingface.co/datasets/google/fleurs) + **Common Voice Corpus 25.0** (Mozilla, March 2025; https://commonvoice.mozilla.org/en/datasets), capped per the paper. Text-only data from the **Goldfish** corpus (Chang et al., 2026).52 53## Limitations54Monolingual (Welsh only). Evaluated on FLEURS / Common Voice test splits; other domains or dialects may differ.55 56## Links & citation57- **Paper:** https://arxiv.org/abs/2609.09554 (Findings of EMNLP 2026)58- **Project page:** https://lemn-lab.github.io/buzz-asr/59- **All models:** https://huggingface.co/BuzzASR60 61```bibtex62@misc{buzzasr2026,63  title         = {BuzzASR: A Swarm of 100+ Monolingual Speech Recognition Models},64  author        = {Shivam Singh and Aditya Yadavalli and Catherine Arnett and Alex Warstadt},65  year          = {2026},66  eprint        = {2609.09554},67  archivePrefix = {arXiv},68  primaryClass  = {cs.CL},69  note          = {Findings of the Association for Computational Linguistics: EMNLP 2026},70  url           = {https://arxiv.org/abs/2609.09554}71}72```73