Speech Engine Comparison
whisper.cpp vs Whisper vs VOSK vs Remote API on Linux
If you are choosing a Linux speech-to-text engine, this page gives a practical side-by-side comparison focused on latency, hardware support, install footprint, privacy boundary, and real desktop usage.
whisper.cpp
- Speed
- Fastest startup + low latency
- Hardware
- CPU + AMD/Intel/NVIDIA GPU
- Accuracy
- High (best overall balance)
- Footprint
- Small models available (~74MB tiny)
- Best for
- Most users who want strong speed + quality
Whisper (OpenAI)
- Speed
- Slower install and startup
- Hardware
- CPU or NVIDIA CUDA
- Accuracy
- High
- Footprint
- Large dependency footprint (~2.3GB)
- Best for
- Users already standardized on PyTorch stack
Faster Whisper
- Speed
- Fast CPU inference (CTranslate2 / INT8)
- Hardware
- CPU
- Accuracy
- High (Whisper-class)
- Footprint
- Optional extra; not in AppImage/Flatpak/Snap
- Best for
- CPU-only machines that want Whisper quality without PyTorch
VOSK
- Speed
- Very fast realtime on low-end systems
- Hardware
- CPU
- Accuracy
- Good for lightweight use
- Footprint
- Very lightweight (~40MB model)
- Best for
- Older hardware and minimal-resource environments
Parakeet
- Speed
- Fast CPU (sherpa-onnx TDT 0.6B)
- Hardware
- CPU
- Accuracy
- High on bundled English / European languages
- Footprint
- Large download (~639MB v3-european default)
- Best for
- English or European dictation without the Whisper catalog picker
Remote API
- Speed
- Depends on server + network latency
- Hardware
- Client CPU + remote Whisper server
- Accuracy
- Depends on remote model
- Footprint
- No local model required
- Best for
- Powerful LAN servers or shared transcription backends
| Engine | Speed | Hardware | Accuracy | Footprint | Best for |
|---|---|---|---|---|---|
| whisper.cpp | Fastest startup + low latency | CPU + AMD/Intel/NVIDIA GPU | High (best overall balance) | Small models available (~74MB tiny) | Most users who want strong speed + quality |
| Whisper (OpenAI) | Slower install and startup | CPU or NVIDIA CUDA | High | Large dependency footprint (~2.3GB) | Users already standardized on PyTorch stack |
| Faster Whisper | Fast CPU inference (CTranslate2 / INT8) | CPU | High (Whisper-class) | Optional extra; not in AppImage/Flatpak/Snap | CPU-only machines that want Whisper quality without PyTorch |
| VOSK | Very fast realtime on low-end systems | CPU | Good for lightweight use | Very lightweight (~40MB model) | Older hardware and minimal-resource environments |
| Parakeet | Fast CPU (sherpa-onnx TDT 0.6B) | CPU | High on bundled English / European languages | Large download (~639MB v3-european default) | English or European dictation without the Whisper catalog picker |
| Remote API | Depends on server + network latency | Client CPU + remote Whisper server | Depends on remote model | No local model required | Powerful LAN servers or shared transcription backends |
Switching Between Engines
You can switch between whisper.cpp, Faster Whisper, Whisper, VOSK, Parakeet, and Remote API from Settings. Speech Model simple setup (v0.17.0) puts language and speed/accuracy first; engine choice lives under Advanced. v0.10.1+ safely stops recognition before switching.
When to pick whisper.cpp
Choose whisper.cpp when you want the best speed-to-accuracy ratio and broad hardware support. It is the default in Vocalinux for a reason. Safe engine switching - v0.10.1+ stops recognition before switching to prevent crashes.
When to pick Whisper
Choose OpenAI Whisper if your environment already depends on PyTorch/CUDA workflows and you prefer that runtime profile.
When to pick Faster Whisper
Choose Faster Whisper for Whisper-class accuracy on CPU without the PyTorch stack. Optional extra; not bundled in AppImage, Flatpak, or Snap.
When to pick VOSK
Choose VOSK on older laptops, low-RAM systems, or lightweight VMs where small model size and minimal overhead matter most.
When to pick Parakeet
Choose Parakeet for English or European dictation on CPU via sherpa-onnx. It uses a bundled language set (v3-european default) instead of the Whisper catalog picker. Optional extra; not in AppImage, Flatpak, or Snap.
When to pick Remote API
Choose Remote API when a trusted server has stronger hardware, larger models, or a shared Whisper backend. Use local engines when your voice data must stay entirely on-device.
Remote setupNext steps
- Install by distro:Ubuntu,Fedora,Arch Linux.
- Use interactive install to detect your hardware and pick the best engine defaults.
- After install, tune model size, VAD sensitivity, or Remote API settings for your preferred latency and accuracy level.