Table with columns: Test Set, Stock Qwen3-ASR-0.6B, This (0.6B-Cantonese-HK), Stock Qwen3-ASR-1.7B, 1.7B-Cantonese-HK v2, whisper-large-v3| Test Set | Stock Qwen3-ASR-0.6B | This (0.6B-Cantonese-HK) | Stock Qwen3-ASR-1.7B | 1.7B-Cantonese-HK v2 | whisper-large-v3 |
|---|
FLEURS (yue_hant_hk) | 6.10% | 5.70% | 4.20% | 3.95% | 20.97% |
Common Voice yue | 6.84% | 5.12% | 5.46% | 3.86% | 36.44% |
Common Voice 23 zh-HK | 9.55% | 6.44% | 7.65% | 5.09% | 41.76% |
| MDCC | 5.45% | 3.29% | 4.25% | 2.49% | 34.52% |
| LegCo (HK Legislative Council) | 8.81% | 7.62% | 7.22% | 7.16% | 32.15% |
| MCE (Cantonese-English Code-Switching) | 15.52% | 6.98% | 11.74% | 5.62% | 24.00% |
Context Hotword Biasing & Anti-Parroting (210-clip stress suite)
Trained with a 75/15/7/3 context-conditioning curriculum (including homophone traps, pure distractors, and ultra-short 1–4 character utterances so the model uses relevant hotwords without parroting unused prompt terms):
Table with columns: Condition, Stock Qwen3-ASR-0.6B CER, This (0.6B-Cantonese-HK) CER, Parrot Rate| Condition | Stock Qwen3-ASR-0.6B CER | This (0.6B-Cantonese-HK) CER | Parrot Rate |
|---|
| Relevant hotwords + homophone distractors | 11.23% | 8.14% | 0.0% |
| 100% irrelevant distractor list | 15.69% | 11.92% | 0.0% |
| Ultra-short audio + long hotword list | 14.38% | 12.50% | 0.0% |
Before scoring, text is converted to Simplified (OpenCC t2s), numbers are normalised and spaces removed. Language is forced to Cantonese.
Only Cantonese and Cantonese–English speech were evaluated; other languages may be worse than the stock model.
import torchfrom qwen_asr import Qwen3ASRModel # pip install qwen-asr asr = Qwen3ASRModel.from_pretrained( "joe2643/Qwen3-ASR-0.6B-Cantonese-HK", dtype=torch.bfloat16, device_map="cuda:0", max_new_tokens=256,) # Standard transcriptionprint(asr.transcribe(audio="sample.wav", language="Cantonese")[0].text) # Optional hotword / domain vocabulary biasing (placed in system prompt under the hood)print(asr.transcribe( audio="command.wav", language="Cantonese", context="閂燈, 開燈, 閂冷氣, 較細聲啲",)[0].text)
Used by
- OraType: free, open-source voice typing for macOS. It ships the GGUF
Q6_K with the mmproj-Q8_0 encoder as its built-in speech model for Cantonese.
Support
If this model is useful to you, consider buying me a lemon tea (請飲杯檸檬茶) to support future open-source Cantonese speech training runs: buymeacoffee.com/joe2643 🍋
Credits
- Fine-tuning, quantization, and GGUF export: @joe2643 (HF)
- Fine-tuning strategy, data collection direction: @sennett-lau