Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
Local Whisper Setup
OpenWhispr supports local speech-to-text processing using whisper.cpp. This keeps your audio completely private—nothing leaves your device.
Quick Start
- Open the Control Panel (right-click tray icon or click the overlay)
- Go to Settings → Speech to Text Processing
- Enable Use Local Whisper
- Select a model (recommended:
base) - Click Save
The first transcription will download the model automatically.
Model Selection
| Model | Size | Speed | Quality | RAM | Best For |
|---|---|---|---|---|---|
| tiny | 75MB | Fastest | Basic | ~1GB | Quick notes |
| base | 142MB | Fast | Good | ~1GB | Recommended |
| small | 466MB | Medium | Better | ~2GB | Professional use |
| medium | 1.5GB | Slow | High | ~5GB | High accuracy |
| large | 3GB | Slowest | Best | ~10GB | Maximum quality |
| turbo | 1.6GB | Fast | High | ~6GB | Fast + accurate |
GPU Acceleration
Local Whisper can run on your GPU for much faster transcription:
- macOS: Metal acceleration is built in — no setup needed on Apple Silicon
- NVIDIA (Windows/Linux): CUDA acceleration; Linux requires a current compatible NVIDIA driver, but not the CUDA Toolkit
- AMD / Intel (Windows/Linux): Vulkan acceleration for Radeon, Arc, and integrated GPUs
- NVIDIA alternative: Vulkan is also supported and is used automatically when its runtime is installed and CUDA fails
GPU runtimes are downloaded on demand with SHA-256-verified checksums. If one fails, OpenWhispr uses the next available backend and finally CPU, so dictation keeps working.
How It Works
OpenWhispr uses whisper.cpp, a high-performance C++ implementation of OpenAI's Whisper model:
- whisper.cpp binary is bundled with the app (or uses system installation as fallback)
- GGML models are downloaded on first use to
~/.cache/openwhispr/whisper-models/ - Audio is processed locally using FFmpeg (bundled with the app)
Requirements
- Disk Space: 75MB–3GB depending on model
- RAM: 1GB–10GB depending on model
- GPU driver: required for GPU acceleration; the CUDA Toolkit is not required
Running From Source
If you're running OpenWhispr locally from a git checkout (not a packaged app), download the whisper.cpp binary for your current platform:
npm run download:whisper-cpp
This puts the binary in resources/bin/. For multi-platform packaging from a single machine, use:
npm run download:whisper-cpp:all
File Locations
| Data | macOS | Windows | Linux |
|---|---|---|---|
| Models | ~/.cache/openwhispr/whisper-models/ |
%USERPROFILE%\.cache\openwhispr\whisper-models\ |
~/.cache/openwhispr/whisper-models/ |
Troubleshooting
"Not Found" Status
- Click Recheck Installation in Control Panel
- Restart the app
- If bundled binary fails, install via package manager:
- macOS:
brew install whisper-cpp - Linux: Build from source at https://github.com/ggml-org/whisper.cpp
- macOS:
Transcription Fails
- Verify microphone permissions
- Try a smaller model (tiny/base)
- Check disk space for model downloads
Slow Performance
- Use smaller models (tiny or base)
- Close resource-intensive apps
- Consider using cloud mode for large files
Privacy Comparison
| Mode | Audio Leaves Device | Internet Required | Cost |
|---|---|---|---|
| Local | No | Only for model download | Free |
| Cloud | Yes (to OpenAI) | Yes | API usage |
Optional local model residency
Turn on Keep model loaded (Settings → Language Models, shown for any feature set to Local) to keep whichever local language model is loaded in memory after idle periods. The default remains the five-minute idle unload.
Residency avoids the next model reload but keeps the model's RAM/VRAM allocated and does not speed up token generation. Explicit shutdown, model changes, and context growth still release/restart the server normally. This option leaves the Local inference timeout, context sizing and cleanup controls unchanged. It also applies when other local-language-model features share the same server. Turn it off to restore the five-minute idle unload.