faster-whisper

Local Whisper via faster-whisper.

Run hyprwhspr setup and select Whisper. On NVIDIA or CPU, setup uses faster-whisper.

Best for: CPU users wanting faster inference than whisper.cpp, or NVIDIA GPU users where VRAM is constrained — INT8 quantization runs large-v3-turbo in ~3.1 GB vs ~6 GB for float16. AMD/Intel GPU users should use Parakeet or whisper.cpp instead (CTranslate2 does not support Vulkan or ROCm).

Built-in Silero VAD strips silence before inference — the most effective mitigation for Whisper’s hallucination loops on longer recordings.

{
    "transcription_backend": "faster-whisper",
    "faster_whisper_model": "large-v3-turbo",   // CUDA; use "base" or "small" for CPU
    "faster_whisper_device": "auto",             // auto | cuda | cpu
    "faster_whisper_compute_type": "auto",       // auto → int8; float16/float32 trade memory for precision
    "faster_whisper_vad_filter": true            // Silero VAD (default: true)
}

Available models

Model	Size (INT8)	Notes
tiny	~75 MB	Fastest
base	~145 MB	Recommended for CPU
small	~484 MB	Better accuracy
medium	~1.5 GB	High accuracy
large-v3	~3.1 GB	Best accuracy (needs GPU)
large-v3-turbo	~1.6 GB	Recommended for CUDA
distil-large-v3	~1.5 GB	Distilled, CPU/GPU balance

Models stored in: ~/.cache/huggingface/hub/

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