Cantonese, written the way Hong Kong speaks it.

Koelony is a team based in Hong Kong, founded to build speech and language models specifically for Hong Kong Cantonese, from open-source releases to the models behind our products. First up is speech recognition, free on Hugging Face.

Qwen3-ASR Cantonese HK

Fine-tuned for Hong Kong Cantonese, written in Traditional Chinese. It handles switching between Cantonese and English the way a Hongkonger does. Not just another Yue ASR.

今gam1朝ziu1喺hai2地dei6鐵tit3站zaam6等dang2咗zo2成seng4粒nap1鐘zung1,個go3schedule完jyun4全cyun4delay晒saai3。
Transcribed by Qwen3-ASR Cantonese HK
  • 1.7B

    Most accurate, for desktops and servers

  • 0.6B

    Small enough to run in a browser

Roadmap

Speech recognition is out. These are next.

  • Cantonese speech recognition

    Qwen3-ASR in 1.7B and 0.6B, open on Hugging Face.

    Released

  • Cantonese text-to-speech

    Hong Kong voices that get all six tones right.

    In training

  • Embeddings & HK-Bench

    Search and retrieval for Cantonese and mixed text, with our own benchmark.

    In training

  • Tone & Jyutping judge

    Scores pronunciation syllable by syllable.

    Research

  • G2P: characters to Jyutping

    Written Cantonese to Jyutping, polyphones included.

    Research

  • Tiny on-device ASR

    For devices too small for the 0.6B model.

    Research

Used in OraType

OraType, a free and open-source voice typing app for macOS, ships our GGUF model as its built-in Cantonese speech engine.

See OraType