BuzzASR/pashto

1---2language: ps3license: mit4library_name: transformers5pipeline_tag: automatic-speech-recognition6base_model: openai/whisper-large-v37tags: [automatic-speech-recognition, whisper, pashto, buzzasr]8datasets: [google/fleurs]9metrics: [cer, wer]10---11 12# BuzzASR — Pashto13 14A monolingual automatic speech recognition model for **Pashto**, 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 **simple fine-tuning (Whisper's tokenizer, ASR fine-tuning only)**.20 21## Results (normalized CER / WER, %)22 23| Test set | CER | WER | Whisper-large-v3 (zero-shot) CER |24|---|---|---|---|25| FLEURS | 19.09 | 47.89 | 35.01 |26| Common Voice 25 | 8.7 | 26.46 | 36.9 |27| Combined | 11.1 | 31.4 | 36.45 |28 29~3.3x CER reduction over Whisper zero-shot on the combined test set.30 31## Usage32 33```python34import torch, torchaudio35from transformers import WhisperForConditionalGeneration, WhisperProcessor36 37model = WhisperForConditionalGeneration.from_pretrained("BuzzASR/pashto", torch_dtype=torch.float16).to("cuda").eval()38proc  = WhisperProcessor.from_pretrained("BuzzASR/pashto")39 40wav, sr = torchaudio.load("audio.wav")           # 16 kHz mono41feats = proc(wav[0], sampling_rate=16000, return_tensors="pt").input_features.to("cuda").half()42ids = model.generate(feats, num_beams=1, no_repeat_ngram_size=3, repetition_penalty=1.2)43print(proc.batch_decode(ids, skip_special_tokens=True)[0])44```45The language/task prompt is baked into the generation config, so no `language=` argument is needed.46 47## Training data48[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).49 50## Limitations51Monolingual (Pashto only). Evaluated on FLEURS / Common Voice test splits; other domains or dialects may differ.52 53## Links & citation54- **Paper:** https://arxiv.org/abs/2609.09554 (Findings of EMNLP 2026)55- **Project page:** https://lemn-lab.github.io/buzz-asr/56- **All models:** https://huggingface.co/BuzzASR57 58```bibtex59@misc{buzzasr2026,60  title         = {BuzzASR: A Swarm of 100+ Monolingual Speech Recognition Models},61  author        = {Shivam Singh and Aditya Yadavalli and Catherine Arnett and Alex Warstadt},62  year          = {2026},63  eprint        = {2609.09554},64  archivePrefix = {arXiv},65  primaryClass  = {cs.CL},66  note          = {Findings of the Association for Computational Linguistics: EMNLP 2026},67  url           = {https://arxiv.org/abs/2609.09554}68}69```70