Qwen3-ASR

Qwen3-ASR, local through a pinned llama.cpp sidecar. Best for Chinese, Japanese and Korean. Thirty languages; 22 Chinese dialects.

Run hyprwhspr setup; choose Qwen3-ASR. It adds no Python packages—only the runtime (17–34 MB) and model pair.

{
    "transcription_backend": "qwen3-asr",
    "qwen3_asr_model": "1.7b-q8_0",   // 1.7b-q8_0 (quality) or 0.6b-q8_0 (smaller)
    "qwen3_asr_device": "auto",       // auto | cpu | vulkan
    "qwen3_asr_timeout": 180,         // sidecar request timeout, seconds (1-600)
    "qwen3_asr_ctx_size": 8192        // llama-server context, tokens (512-65536); null = llama.cpp default
}

auto keeps an installed runtime. Otherwise: Vulkan when found, CPU when not. Vulkan covers NVIDIA, AMD and Intel; there is no Linux CUDA build.

qwen3_asr_ctx_size sizes the KV cache, ~112 KiB per token on the 1.7B. 8192 is ~0.9 GiB and covers ~5 minutes of audio; requests carry at most 2. null restores llama.cpp’s 32000, ~3.5 GiB.

Long audio splits at pauses, then joins again. Set language when you can; without it, the first useful segment guides the rest.

Available models

Model	Size	Notes
1.7b-q8_0	~2.4 GB	Recommended · best quality
0.6b-q8_0	~1.0 GB	Smaller · faster · less accurate

Models stored in: ~/.local/share/hyprwhspr/qwen3-asr/models/

Languages

Chinese, Cantonese, English, Japanese, Korean, Arabic, German, French, Spanish, Portuguese, Italian, Russian, Dutch, Polish, Turkish, Thai, Vietnamese, Hindi, Indonesian, Malay and more.

Leave language unset to detect it. Set an ISO code to hold it steady.

Note: No streaming, timestamps, alignment or prompts. whisper_prompt_* does not apply.