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%.
Model size

Lower is better.

Character error rate (1.7B). Lower is better.
Test setKoelony fine-tuneStock Qwen3-ASRWhisper large-v3
FLEURS yue_hant_hk3.89%4.20%20.97%
Common Voice yue3.94%5.46%36.44%
Common Voice 23 zh-HK5.02%7.65%41.76%
MDCC2.47%4.25%34.52%
LegCo7.16%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 repoFormatSizeRuntime
Qwen3-ASR-1.7B-Cantonese-HKBF16 safetensors4.08 GBtransformers, qwen-asr, vLLM, ComfyUI
Qwen3-ASR-1.7B-Cantonese-HK-GGUFGGUF Q4_K_M / Q6_K / Q8_01.1–1.8 GBllama.cpp, OraType
Qwen3-ASR-1.7B-Cantonese-HK-FP8FP8 W8A82.67 GBvLLM
Qwen3-ASR-1.7B-Cantonese-HK-NVFP4NVFP4A162.05 GBvLLM
Qwen3-ASR-0.6B-Cantonese-HKBF16 safetensors1.56 GBtransformers, qwen-asr, vLLM
Qwen3-ASR-0.6B-Cantonese-HK-GGUFGGUF Q4_K_M / Q6_K / Q8_0397–639 MBllama.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.

Open the demo

The first visit downloads about 1.2 GB. Works best in desktop Chrome or Edge.