Ghost Pepper

Ghost Pepper is a hold-to-talk dictation tool for macOS: hold Control, speak, release, and the transcribed text pastes directly into whatever field you were focused on. It also handles meeting transcription, recording calls and producing notes, a transcript, and an AI-generated summary saved as markdown, all processed locally.

Everything runs on-device via Apple Silicon (M1 or newer, macOS 14.0+) using local voice models for speech-to-text across 50+ languages, plus a local LLM that cleans up filler words and self-corrections from the raw transcript before it’s inserted. It lives in the menu bar with no dock icon, can launch at login, and lets you customize the cleanup prompt, microphone, and which features are active.

The README describes the app as free and open source with an MIT badge, but no LICENSE file is present in the repository at the time of review — meaning the actual legal terms default to standard copyright (all rights reserved) despite the stated intent, until a license file is formally added. The source is publicly viewable and the compiled app is distributed as a free download.

What You Get

• Hold-Control-to-talk dictation that transcribes and pastes directly into any text field
• Meeting transcription with notes, full transcript, and an AI-generated summary saved as markdown
• Fully on-device processing via Apple Silicon, with no cloud API calls or uploaded audio
• A menu bar app with launch-at-login, customizable cleanup prompts, and mic selection

Common Use Cases

• Dictating text into any macOS application without sending audio to a cloud speech-to-text service
• Recording and transcribing meetings locally, with an AI-generated summary for later reference
• Working in privacy-sensitive contexts where audio can’t leave the device for compliance or confidentiality reasons
• Dictating in one of 50+ supported languages without depending on internet connectivity

Under The Hood

Architecture Ghost Pepper is a native Swift macOS menu bar app that pairs a local speech-to-text model with a local LLM cleanup pass — the two-stage pipeline (raw transcription, then LLM-based filler-word and self-correction cleanup) is what lets it produce cleaner output than a raw speech-to-text model alone, without any of that processing leaving the device. The meeting-transcription feature layers recording, transcription, and summarization together, all writing to markdown output rather than a proprietary format.

Tech Stack Swift, targeting macOS 14.0+ on Apple Silicon (M1 and newer) specifically, using on-device voice models for speech recognition and a local LLM for text cleanup — no server component or cloud dependency of any kind.

Code Quality Very active, consistently maintained commit history reflects ongoing development, though as a young, single-developer project it hasn’t yet accumulated the maturity signals (broad test coverage, long release history) of longer-running tools; notably, the repository lacks a LICENSE file despite the README’s MIT badge, worth resolving before broader adoption.

What Makes It Unique Many dictation tools rely on cloud speech APIs even when marketed as “private”; Ghost Pepper’s entire pipeline — transcription and LLM-based cleanup — runs on-device via Apple Silicon, which is a meaningfully stronger privacy guarantee than tools that merely encrypt data in transit to a server.

Self-Hosting

Licensing Model The README displays an MIT license badge and describes the project as “Free & open source,” but no LICENSE file exists in the repository as of this review. Legally, this means the code defaults to standard copyright (all rights reserved) despite the stated open-source intent — treat this as source-available rather than a formally licensed open-source project until a LICENSE file is added.

Self-Hosting Restrictions Not applicable; it’s a local macOS app with no hosted service.

License Key Required No; the compiled app is distributed as a free download.