whisper.cpp

Local Whisper via pywhispercpp.

Run hyprwhspr setup and select Whisper. On AMD/Intel, setup uses whisper.cpp with Vulkan.

For whisper.cpp on CPU or NVIDIA: hyprwhspr setup auto --backend cpu or --backend nvidia.

On x86-64, setup fetches a pre-built wheel: CUDA 12, or Vulkan on glibc 2.41+ (Arch, Fedora 42+, Debian 13). Otherwise it builds, and a Vulkan build needs:

• Arch: setup installs it
• Debian/Ubuntu: libvulkan-dev glslc spirv-headers
• Fedora: vulkan-headers vulkan-loader-devel glslc spirv-headers-devel
• openSUSE: vulkan-devel shaderc spirv-headers

If the GPU build fails, you get CPU, and the service says so at every start. Install what’s missing, re-run setup, reinstall the backend.

Best for: modern NVIDIA cards or discrete AMD/Intel (via Vulkan) — extremely fast on GPU with large-v3 or large-v3-turbo.

Available models

Models stored in: ~/.local/share/pywhispercpp/models/

Model	Size	Notes
tiny / tiny.en	~75 MB	Fastest
base / base.en	~148 MB	Recommended (default)
small / small.en	~488 MB	Better accuracy
medium / medium.en	~1.5 GB	High accuracy
large-v3	~2.9 GB	Best accuracy, requires GPU
large-v3-turbo	~1.6 GB	Fast + accurate, requires GPU

GPU required: large-v3 and large-v3-turbo require GPU acceleration for reasonable speed.

Download a specific model by name:

hyprwhspr model download base
hyprwhspr model download small.en

Set model in config (pywhispercpp only — faster-whisper uses faster_whisper_model):

{
    "model": "small.en",  // .en = English-only; omit suffix for multilingual
    // "threads": 6       // optional; omit for auto = min(8, CPU count)
}

Voice activity detection

Optional native Silero VAD strips silence before inference — the same hallucination mitigation faster-whisper ships, off by default here because it needs an extra ~1 MB model (ggml-silero-v5.1.2.bin, auto-downloaded to the models directory on first use):

{
    "pywhispercpp_use_vad": true   // default: false
}

If the download fails (e.g. offline), the service logs a warning and continues without VAD.

Language, prompts, translation and decoding: Language and prompts.