BuzzASR — Korean

A monolingual automatic speech recognition model for Korean, fine-tuned from openai/whisper-large-v3. Part of BuzzASR, a suite of 102 language-specialized ASR models (paper: arXiv:2609.09554, Findings of EMNLP 2026).

This model uses simple fine-tuning (Whisper's tokenizer, ASR fine-tuning only).

🏆 State-of-the-art (open-source). On the combined FLEURS + Common Voice test set, this model achieves the lowest CER of every open system we compare against: Whisper-large-v3, Omnilingual 1B/7B, MMS, Qwen3-ASR, and Cohere Transcribe.

Results (normalized CER / WER, %)

Test set	CER	WER	Whisper-large-v3 (zero-shot) CER
FLEURS	5.16	13.22	5.31
Common Voice 25	4.24	14.11	6.01
Combined	4.61	13.74	5.72

~1.2x CER reduction over Whisper zero-shot on the combined test set.

How the scores above were computed

CER/WER are computed with jiwer 3.0.4 after applying the same normalization to references and hypotheses: lowercase, remove all characters except letters, combining marks, numbers, whitespace, apostrophes and hyphens, then collapse whitespace. This removes parentheses but keeps the words inside them, so FLEURS scores include parenthesised words that are not spoken in the audio. Spacing differences are also counted as errors.

Recommended decoding settings

An independent evaluation by Kwan-u Choi with kocer (v0.5.0), a Korean-specific CER that accounts for notation and spacing differences, found that beam search (num_beams=5) gives the lowest CER on all six Korean test sets evaluated, while the settings this card previously recommended (num_beams=1, no_repeat_ngram_size=3, repetition_penalty=1.2) gave the highest. We now recommend num_beams=5. Full results are in the kocer repository; scores are from kocer 0.5.0 and may differ in other versions. Thanks to Kwan-u for the evaluation:)

Usage

import torch, torchaudio
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model = WhisperForConditionalGeneration.from_pretrained("BuzzASR/korean", torch_dtype=torch.float16).to("cuda").eval()
proc  = WhisperProcessor.from_pretrained("BuzzASR/korean")

wav, sr = torchaudio.load("audio.wav")           # 16 kHz mono
feats = proc(wav[0], sampling_rate=16000, return_tensors="pt").input_features.to("cuda").half()
ids = model.generate(feats, num_beams=5)         # recommended; see "Recommended decoding settings"
print(proc.batch_decode(ids, skip_special_tokens=True)[0])

The language/task prompt is baked into the generation config, so no language= argument is needed.

Training data

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).

Limitations

Monolingual (Korean only). Evaluated on FLEURS / Common Voice test splits; other domains or dialects may differ.

Links & citation

• Paper: https://arxiv.org/abs/2609.09554 (Findings of EMNLP 2026)
• Project page: https://lemn-lab.github.io/buzz-asr/
• All models: https://huggingface.co/BuzzASR

@misc{buzzasr2026,
  title         = {BuzzASR: A Swarm of 100+ Monolingual Speech Recognition Models},
  author        = {Shivam Singh and Aditya Yadavalli and Catherine Arnett and Alex Warstadt},
  year          = {2026},
  eprint        = {2609.09554},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  note          = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  url           = {https://arxiv.org/abs/2609.09554}
}