BuzzASR / cantonese

BuzzASR — Cantonese

A monolingual automatic speech recognition model for Cantonese , fine-tuned from openai/whisper-large-v3 . Part of BuzzASR , a suite of 102 language-specialized ASR models (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, %)

~2.8x CER reduction over Whisper zero-shot on the combined test set.

Usage

import torch, torchaudio
from transformers import WhisperForConditionalGeneration, WhisperProcessor

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

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=1, no_repeat_ngram_size=3, repetition_penalty=1.2)
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 (Cantonese only). Evaluated on FLEURS / Common Voice test splits; other domains or dialects may differ.

Citation

Project page: https://lemn-lab.github.io/buzzasr-docs/