matthartman/ghost-pepper
100% private on-device voice models for speech-to-text and meeting transcription on macOS
What it solves
Ghost Pepper is a privacy-focused voice-to-text tool for macOS that allows users to transcribe speech and record meetings without sending any audio or text data to cloud APIs. It eliminates the dependency on external servers for transcription and AI-driven text cleanup, ensuring that sensitive conversations remain entirely on the local machine.
How it works
The app runs entirely on Apple Silicon (M1+) using a suite of open-source models cached locally. It utilizes WhisperKit, FluidAudio, and MLX Audio for speech-to-text (ASR) across various models like Whisper and Qwen3-ASR, and LLM.swift to run local LLMs (such as Qwen 3.5) for removing filler words and generating meeting summaries. Users can trigger transcription via a global hotkey (Control key) or record meetings to save as markdown files.
Who it’s for
Mac users who prioritize privacy and security, professionals who record meetings and need automated summaries, and anyone who want a voice-to-text tool that works offline and supports multiple languages.
Highlights
- 100% Local Execution: All transcription, cleanup, and summarization happen on-device via Apple Silicon.
- Multi-model Support: Offers a variety of speech models for different needs, including low-latency streaming and multilingual support for 50+ languages.
- Smart Cleanup: Uses a local LLM to automatically remove filler words and handle self-corrections in transcribed text.
- Meeting Transcription: Records calls and generates notes, transcripts, and AI summaries saved as local markdown files.
- Menu Bar Integration: Operates as a lightweight menu bar app that can launch at login.
Related
- Project
- Project
- Project
- Project
- Project