Starling – Local, Free, Fast, Speech to Text (github.com)

Starling is a lightweight macOS menu-bar app that lets you dictate anywhere with a global hotkey (default ⌃⌥⌘J) and instantly pastes transcribed text at your cursor. It performs local speech-to-text on-device using NVIDIA Parakeet v3 exposed through a FluidAudio Core ML wrapper, leveraging your Mac’s Neural Engine for fast, privacy-preserving transcription. Voice activity detection (VAD) automatically stops when you finish speaking, and the app simulates a paste (or copies to clipboard) without stealing focus. The first run downloads a ~2.5 GB Parakeet Core ML model to ~/Library/Caches/FluidAudio; afterwards no audio or text leaves your machine. For the AI/ML community this demonstrates practical on-device ASR with low latency and strong privacy guarantees—useful for apps, researchers and privacy-minded users who want offline transcription without cloud dependencies. Technical caveats: Starling is optimized for English (Parakeet supports ~25 languages), it transcribes after you stop speaking (no live streaming yet), and requires macOS 14+, Apple Silicon or Intel Macs with Neural Engine support plus Accessibility and microphone permissions. The project is open-source (GitHub) with build instructions (Xcode 15, Swift 6) and a roadmap for streaming partial results, multi-language UI and model management—making it a useful reference for local Core ML ASR deployments.