Choosing a backend

Quick pick by hardware:

NVIDIA GPU → Cohere Transcribe
AMD / Intel GPU → Parakeet.cpp · whisper.cpp (Vulkan) for 99 languages
CPU only → Parakeet or faster-whisper
ARM64 → Parakeet
Chinese, Japanese or Korean → Qwen3-ASR
No local setup → REST API

Model	Engine	Runs on	Arch	Speed	Accuracy	Memory	Languages
Parakeet v3	ONNX	CPU · NVIDIA	x64 · ARM64	●●●	●●○	1 GB	25 European
Orukeet †	ONNX	CPU · NVIDIA	x64 · ARM64	●●●	●●○	1 GB	25 European
Parakeet v3 †	Parakeet.cpp	CPU · Vulkan	x64 · ARM64	●●●	●●○	0.9 GB	25 European
Whisper turbo	whisper.cpp · faster-whisper	CPU · NVIDIA · Vulkan	x64 · ARM64	●●○	●●○	1.6 GB	99
Cohere Transcribe	PyTorch	CPU · NVIDIA	x64 · ARM64	●●●	●●●	4 GB	14
Qwen3-ASR 1.7B †	llama.cpp	CPU · Vulkan	x64	●●○	●●○ · ●●● CJK	2.4 GB	30

Speed on each model’s best hardware. Accuracy from the Open ASR Leaderboard. Memory is RAM or VRAM, roughly. ARM64 runs on CPU. Whisper ships smaller models: faster-whisper, whisper.cpp. † Experimental.

Cloud or your own server: REST API and Realtime WebSocket. Speed and accuracy follow the provider.

hyprwhspr setup auto picks Whisper for your hardware: faster-whisper on NVIDIA or CPU, whisper.cpp on AMD/Intel.

Model commands

hyprwhspr model commands route automatically to the configured local backend. For cloud backends (rest-api and realtime-ws), model operations are not applicable and exit nonzero.

hyprwhspr model status            # Check if model is downloaded/cached
hyprwhspr model list              # Show model info for active backend
hyprwhspr model download [model]  # Download or re-download model
hyprwhspr model unload            # Free its memory; the service keeps running
hyprwhspr model reload            # Reload the model after an unload

Models are downloaded automatically during hyprwhspr setup; use model download to re-download if needed. To free memory, see Memory management.