Ghost Pepper: Free open-source transcription that stays on your Mac

Matt Hartman built Ghost Pepper, a macOS app that does voice dictation and meeting transcription without sending anything to the cloud. It runs entirely on Apple Silicon using WhisperKit for speech-to-text and Qwen LLMs for text cleanup and summaries. Hold Control, talk, release. Your words appear in whatever text field you're using. Record a meeting and get notes, a transcript, and an AI summary saved as local Markdown. The app requires macOS 14.0 or later and an M1 chip or newer.

Ghost Pepper enters a space dominated by MacWhisper and Superwhisper. MacWhisper runs a freemium model, gating advanced features and larger models behind a paid subscription. Ghost Pepper is free and open source under an MIT license. It also uses WhisperKit, Apple's official Swift implementation of Whisper, rather than the whisper.cpp C++ port that most competitors rely on. That could mean better optimization for Apple's hardware, though real-world benchmarks would tell the full story.

The project hit #1 on Hacker News. Commenters traded notes on how local speech recognition has actually gotten good. They compared modern transformer models like Whisper and Parakeet against older approaches like Google's WebSpeech API and legacy Windows speech tools. Several pointed out that while Google Pixels have done offline transcription for years, current open-source models handle noise and accents much better. The discussion ranged from OS/2 voice interfaces to Microsoft OneNote, with a clear sense that local AI is finally becoming practical.

What sets Ghost Pepper apart is bundling on-device LLMs for cleanup and summarization right out of the box. It removes filler words and fixes self-corrections automatically. Competitors offer AI features, but they often require separate cloud subscriptions or don't integrate text processing as tightly into a fully local workflow. With models from 75MB to 2.8GB depending on your accuracy needs, and support for over 50 languages, Ghost Pepper makes a strong case that transcription doesn't need to phone home. Some developers switched to Qwen.