Ghost Pepper 🌶️

Hold the Control key, speak, and release. Ghost Pepper transcribes your speech and pastes it into whatever text field is active, whether that's a chat box, an email, or a code editor. It's a single gesture with no window to manage, which is what makes it fast enough to use all day.

Every model runs locally via Apple Silicon. The app downloads open-source models once, caches them, and never sends audio anywhere. That means no account, no upload, and no terms-of-service question about who gets to hear your recordings. Compare that to cloud dictation tools, which route every sentence you speak through a remote server before the text shows up in your document, and the privacy gap is easy to see.

Ghost Pepper records calls and produces notes, a full transcript, and an AI-generated summary, all saved as a markdown file. The summary turns a long call into something you can skim. It's the difference between an hour of audio and a page you can actually use.

A local LLM cleans up what the speech model hears. It removes filler words and handles self-corrections, so "we should, uh, ship it Monday, no wait, Tuesday" comes out as "we should ship it Tuesday." You can edit the cleanup prompt yourself if the default style doesn't match how you write.

You pick the trade-off between speed, size, and language coverage. Whisper tiny.en is fastest and English-only at about 75 MB, while Qwen3-ASR handles 50+ languages at around 900 MB. Parakeet v3 covers 25 languages, and Nemotron streaming models target low-latency work. If you want open-source transcription you can inspect and swap out yourself, this is where you do it.

The app lives in your menu bar with no dock icon and can launch at login. You choose your mic, toggle features on and off, and swap models from the same menu.