# Ghost Pepper
**100% private** on-device voice models for speech-to-text and meeting transcription on macOS. No cloud APIs, no data leaves your machine.
macOS 14.0+ · Apple Silicon (M1+) · Free & open source
[](https://github.com/matthartman/ghost-pepper)
[](LICENSE)
## Features
- **Hold Control to talk** — release to transcribe and paste into any text field
- **Meeting transcription** — record calls with notes, transcript, and AI-generated summaries saved as markdown
- **Runs entirely on your Mac** — models run locally via Apple Silicon, nothing is sent anywhere
- **Smart cleanup** — local LLM removes filler words and handles self-corrections
- **Menu bar app** — lives in your menu bar, no dock icon, launches at login
- **Customizable** — edit the cleanup prompt, pick your mic, toggle features on/off
## How it works
Ghost Pepper uses open-source models that run entirely on your Mac. Models download automatically and are cached locally.
### Speech models
| Model | Size | Best for |
|---|---|---|
| Whisper tiny.en | ~75 MB | Fastest, English only |
| **Whisper small.en** (default) | ~466 MB | Best accuracy, English only |
| Whisper small (multilingual) | ~466 MB | Multi-language support |
| Parakeet v3 (25 languages) | ~1.4 GB | Multi-language via [FluidAudio](https://github.com/FluidInference/FluidAudio) |
| Qwen3-ASR 0.6B int8 (50+ languages) | ~900 MB | Highest multilingual quality, macOS 15+ required |
| Nemotron Speech Streaming 0.6B 8-bit | ~633 MB | Low-latency English streaming via [MLX Audio](https://github.com/Blaizzy/mlx-audio-swift) |
| Nemotron 3.5 ASR Streaming 0.6B 8-bit | ~721 MB | Low-latency multilingual streaming via [MLX Audio](https://github.com/Blaizzy/mlx-audio-swift) |
### Cleanup models
| Model | Size | Speed |
|---|---|---|
| **Qwen 3.5 0.8B** (default) | ~535 MB | Very fast (~1-2s) |
| Qwen 3.5 2B | ~1.3 GB | Fast (~4-5s) |
| Qwen 3.5 4B | ~2.8 GB | Full quality (~5-7s) |
Speech models powered by [WhisperKit](https://github.com/argmaxinc/WhisperKit), [FluidAudio](https://github.com/FluidInference/FluidAudio), and [MLX Audio](https://github.com/Blaizzy/mlx-audio-swift). Cleanup models powered by [LLM.swift](https://github.com/eastriverlee/LLM.swift). All models served by [Hugging Face](https://huggingface.co/).
## Getting started
**Download the app:**
1. Download [GhostPepper.dmg](https://github.com/matthartman/ghost-pepper/releases/latest/download/GhostPepper.dmg)
2. Open the DMG, drag Ghost Pepper to Applications
3. Grant Microphone and Accessibility permissions when prompted
4. Hold Control and speak
> **"Apple could not verify" warning?** On macOS Sequoia, you may see a Gatekeeper warning the first time you open the app. Go to **System Settings > Privacy & Security**, scroll down, and click **Open Anyway** next to the Ghost Pepper message. Click **Confirm** in the popup. You only need to do this once.
**Build from source:**
1. Clone the repo
2. Open `GhostPepper.xcodeproj` in Xcode
3. Build and run (Cmd+R)
## Permissions
| Permission | Why |
|---|---|
| Microphone | Record your voice |
| Accessibility | Global hotkey and paste via simulated keystrokes |
## Privacy audit
Every core feature runs 100% on your Mac — verified by AI code review. No trust required, just point Claude at the repo and ask.
| Feature | Status | What was checked |
|---|---|---|
| Speech-to-text | :white_check_mark: Local | WhisperKit/FluidAudio/MLX Audio inference, no audio sent anywhere |
| Text cleanup | :white_check_mark: Local | Qwen LLM runs on-device via LLM.swift |
| Audio recording | :white_check_mark: Local | AVAudioEngine + ScreenCaptureKit, no streaming |
| Meeting transcription & storage | :white_check_mark: Local | Chunked transcription, markdown files on disk |
| Summary generation | :white_check_mark: Local | Local LLM summarization, no cloud API |
| OCR & screen capture | :white_check_mark: Local | Apple Vision framework, on-device |
| File storage | :white_check_mark: Local | Markdown to local filesystem, no cloud sync |
| Analytics & telemetry | :white_check_mark: Local | Local-only usage counters (UserDefaults) power the in-app Usage report panel. No Firebase, Mixpanel, Sentry, or any tracking SDK. |
**Optional cloud features** (disabled by default, require your own API keys): Zo AI chat, Trello integration, Granola meeting import. Model downloads are one-time from Hugging Face.
> **Verify it yourself:** run `cat PRIVACY_AUDIT.md` in Claude Code and ask it to review the codebase against the audit prompt. The [full audit](PRIVACY_AUDIT.md) includes the exact prompt and detailed file-level results.
## Good to know
- **Launch at login** is enabled by default on first run. You can toggle it off in Settings.
- **Everything stays local** — transcription history and recordings are stored on your Mac only. Nothing is sent to the cloud. You can clear history anytime in Settings.
## Acknowledgments
Built with [WhisperKit](https://github.com/argmaxinc/WhisperKit), [LLM.swift](https://github.com/eastriverlee/LLM.swift), [Hugging Face](https://huggingface.co/), and [Sparkle](https://sparkle-project.org/).
## License
MIT
## Why "Ghost Pepper"?
All models run locally, no private data leaves your computer. And it's spicy to offer something for free that other apps have raised $80M to build.
## Enterprise / managed devices
Ghost Pepper requires Accessibility permission, which normally needs admin access to grant. On managed devices, IT admins can pre-approve this via an MDM profile (Jamf, Kandji, Mosaic, etc.) using a Privacy Preferences Policy Control (PPPC) payload:
| Field | Value |
|---|---|
| Bundle ID | Use the signed app's bundle identifier |
| Team ID | Use the signing team's Apple Developer team ID |
| Permission | Accessibility (`com.apple.security.accessibility`) |
---
Questions or deployment issues: open an issue or use your team's internal support channel.