More love for CJK brothers and sisters - Qwen3-ASR, by all accounts the strongest for Asian performance. 🇯🇵 🇨🇳 🇰🇷
New features
Qwen3-ASR
- Best WER for Chinese, Japanese, and Korea; many Chinese dialects
- Pinned
llama.cppsidecar - Vulkan on NVIDIA, AMD and Intel
- CPU fallback
- 1.7B quality model or lighter 0.6B model
- Long recordings split at pauses, then join again
- Full setup, status, model and unload support
- No extra Python inference packages
Still experimental. No streaming, timestamps or prompt conditioning — yet.
Also inside: refreshed website dependencies and quieter docs.
Choose it from hyprwhspr setup, or run hyprwhspr setup auto --backend qwen3-asr.
Any issues, please let me know! (In English please 😀 )
See Qwen3-ASR configuration for models, hardware support and settings.