Show HN: Local Push-to-Transcribe (github.com)

Transcrybe is a lightweight, fully offline push-to-transcribe app for macOS that turns a single Function key press into near-instant speech-to-text input. It runs entirely locally (no cloud, tracking, or subscription), inserts transcriptions directly into the focused input field, and claims sub‑second results by keeping the chosen speech model resident in RAM. The app is free, open source, has a minimal UI (menu bar settings), and requests only microphone and accessibility permissions to operate—Windows support is in early development and contributions are welcome. Technically, Transcrybe is built on whisper.cpp, a high-performance C++ reimplementation of OpenAI’s Whisper, which enables efficient on-device inference. You choose between model sizes that trade accuracy for memory: roughly 40 MB, 140 MB, or 440 MB kept in memory for fast responses. For the AI/ML community this showcases practical edge ASR: privacy-preserving, low-latency transcription without cloud costs, and a reproducible, open stack that can accelerate experimentation and deployment of on-device speech models. The main constraints are model memory footprint and current macOS-only availability, but the project demonstrates how optimized implementations like whisper.cpp make real-time, private speech interfaces feasible on consumer hardware.