Qwen3-ASR for Hong Kong Cantonese
It writes the characters Hong Kong people use when they speak (係, 喺, 嘅, 咗, 㗎), keeps English words in English, and does not drift into Simplified or Mainland written forms. Open weights under Apache-2.0.
Two sizes
1.7B
Most accurate v2.1
For desktops and servers. Handles short voice commands, and takes a hotword list without repeating it back: 0% parroting in our 210-clip stress test.
0.6B
About a third of the size v1
Trained on the same voice commands and hotwords. Small enough to run in a browser tab.
Character error rate
- On FLEURS Cantonese the 1.7B model gets 3.89% of characters wrong; Whisper large-v3 gets 20.97% wrong.
- On Cantonese–English code-switching (MCE), errors drop from 11.74% with the stock Qwen3-ASR 1.7B to 5.52%.
- The 0.6B model more than halves code-switching errors too, from 15.52% to 6.98%.
Lower is better.
| Test set | Koelony fine-tune | Stock Qwen3-ASR | Whisper large-v3 |
|---|---|---|---|
| FLEURS yue_hant_hk | 3.89% | 4.20% | 20.97% |
| Common Voice yue | 3.94% | 5.46% | 36.44% |
| Common Voice 23 zh-HK | 5.02% | 7.65% | 41.76% |
| MDCC | 2.47% | 4.25% | 34.52% |
| LegCo | 7.16% | 7.22% | 32.15% |
| MCE (code-switching) | 5.52% | 11.74% | 24.00% |
Greedy decoding, BF16 transformers, language forced to Cantonese. Text is converted to Simplified, numbers normalised and spaces removed before scoring. Full method in the repo.
Weights and quantizations
| Hugging Face repo | Format | Size | Runtime |
|---|---|---|---|
| Qwen3-ASR-1.7B-Cantonese-HK | BF16 safetensors | 4.08 GB | transformers, qwen-asr, vLLM, ComfyUI |
| Qwen3-ASR-1.7B-Cantonese-HK-GGUF | GGUF Q4_K_M / Q6_K / Q8_0 | 1.1–1.8 GB | llama.cpp, OraType |
| Qwen3-ASR-1.7B-Cantonese-HK-FP8 | FP8 W8A8 | 2.67 GB | vLLM |
| Qwen3-ASR-1.7B-Cantonese-HK-NVFP4 | NVFP4A16 | 2.05 GB | vLLM |
| Qwen3-ASR-0.6B-Cantonese-HK | BF16 safetensors | 1.56 GB | transformers, qwen-asr, vLLM |
| Qwen3-ASR-0.6B-Cantonese-HK-GGUF | GGUF Q4_K_M / Q6_K / Q8_0 | 397–639 MB | llama.cpp, edge devices |
Runs entirely in your browser.
Try the 0.6B model without installing anything. Record from your mic or drop in an audio file; nothing is uploaded.
The first visit downloads about 1.2 GB. Works best in desktop Chrome or Edge.