About joe2643
building open-source Hong Kong Cantonese speech AI
Hi, I'm Joe — an indie developer building open-source speech AI and local-first tools, with a focus on real Hong Kong Cantonese.
Most speech-to-text models either struggle with Cantonese or forcibly rewrite spoken Cantonese into formal Standard Written Chinese, mangling natural Cantonese–English code-switching along the way. To fix that, I fine-tuned and open-sourced Qwen3-ASR-1.7B-Cantonese-HK (along with GGUF, FP8, and NVFP4 quantizations for llama.cpp, vLLM, and on-device dictation apps like OraType) — so local ASR can finally transcribe authentic Hong Kong Traditional Chinese characters (係、喺、嘅、咗、㗎) and mixed English words the way we actually speak.
Every coffee goes directly toward GPU compute for training, quantization, and releasing lighter on-device Cantonese speech models (like upcoming sub-1B variants and Cantonese TTS) as open weights for the community.
---
我係一個專注整本地優先（local-first）開源工具同香港廣東話語音 AI 嘅獨立開發者。
市面上好多語音識別一係聽唔明廣東話，一係就強行將我哋講嘅口語洗成書面語，遇到中英夾雜更加炒晒車。所以我微調同開源咗 Qwen3-ASR-1.7B-Cantonese-HK（連埋 GGUF / FP8 / NVFP4 量化版，支援 llama.cpp 本地離線跑同埋 Mac 語音輸入），原汁原味保留返香港人平時講嘢用開嘅正字（係、喺、嘅、咗、㗎）同英文詞彙。
你請嘅每一杯咖啡／凍檸茶，都會直接用嚟補貼訓練同評測模型嘅 GPU 算力開支，等我可以繼續開源更輕量、更準嘅香港廣東話 ASR 同 TTS 模型畀大家免費自架同離線用。多謝支持！
Most speech-to-text models either struggle with Cantonese or forcibly rewrite spoken Cantonese into formal Standard Written Chinese, mangling natural Cantonese–English code-switching along the way. To fix that, I fine-tuned and open-sourced Qwen3-ASR-1.7B-Cantonese-HK (along with GGUF, FP8, and NVFP4 quantizations for llama.cpp, vLLM, and on-device dictation apps like OraType) — so local ASR can finally transcribe authentic Hong Kong Traditional Chinese characters (係、喺、嘅、咗、㗎) and mixed English words the way we actually speak.
Every coffee goes directly toward GPU compute for training, quantization, and releasing lighter on-device Cantonese speech models (like upcoming sub-1B variants and Cantonese TTS) as open weights for the community.
---
我係一個專注整本地優先（local-first）開源工具同香港廣東話語音 AI 嘅獨立開發者。
市面上好多語音識別一係聽唔明廣東話，一係就強行將我哋講嘅口語洗成書面語，遇到中英夾雜更加炒晒車。所以我微調同開源咗 Qwen3-ASR-1.7B-Cantonese-HK（連埋 GGUF / FP8 / NVFP4 量化版，支援 llama.cpp 本地離線跑同埋 Mac 語音輸入），原汁原味保留返香港人平時講嘢用開嘅正字（係、喺、嘅、咗、㗎）同英文詞彙。
你請嘅每一杯咖啡／凍檸茶，都會直接用嚟補貼訓練同評測模型嘅 GPU 算力開支，等我可以繼續開源更輕量、更準嘅香港廣東話 ASR 同 TTS 模型畀大家免費自架同離線用。多謝支持！
Recent supporters