Developer Matt Hartman recently released a macOS app called Ghost Pepper, designed to provide users with a completely localized speech-to-text experience. The app works by holding down the Control key to record; once released, it automatically transcribes the audio and pastes it into any active text field.
Unlike many voice input tools on the market that rely on cloud-based APIs, Ghost Pepper processes everything locally on the device. By harnessing the power of Apple Silicon, the app ensures that user data never leaves the computer.
Local Models and Intelligent Optimization
At its core, Ghost Pepper uses WhisperKit for speech recognition, combined with a local large language model based on LLM.swift for audio refinement. This "smart cleanup" feature automatically removes filler words and handles self-corrections, significantly improving the quality of the output.
According to the project documentation, the app defaults to the Whisper small.en model and the Qwen 3.5 0.8B model to strike a balance between responsiveness and accuracy. Users can also opt for larger models, such as the Parakeet v3 multilingual model, or customize the output style by adjusting the cleanup prompts.
The app runs as a menu bar utility, keeping the Dock clutter-free, and supports automatic startup. Regarding privacy, the developer guarantees that the software does not write transcriptions to the disk; all debug logs reside strictly in memory and vanish once the app is closed.
Ghost Pepper is currently open-source on GitHub under the MIT license. It requires macOS 14.0 or later and must be run on a Mac with Apple Silicon. Because the app requires simulated keystrokes and audio recording capabilities, users must manually grant the necessary Accessibility and Microphone permissions upon installation.
For enterprise environments, IT administrators can pre-approve these permissions via MDM configuration profiles to ensure seamless deployment on managed devices. In the project introduction, Hartman noted that "Ghost Pepper provides a free, localized alternative that challenges commercial transcription products currently backed by tens of millions of dollars in funding."
