Sourdine – open-source macOS app for meeting transcription with 100% local AI

  I built a macOS app that transcribes meetings in real-time and generates notes (summary, action items, key points) using AI that runs entirely on your machine. No cloud, no API keys, no
  subscription.

  It captures both your microphone and system audio (Teams, Meet, Zoom) simultaneously using a native Rust module with ScreenCaptureKit, then runs Parakeet TDT (NVIDIA, 640MB) for speech-to-text and
  Mistral 7B for summarization — all local via sherpa-onnx and node-llama-cpp.

  Currently French only — multi-language support is on the roadmap. Pull requests are welcome if you want to help add other languages or contribute in any way.

  Stack: Electron 34, Angular 21, NestJS 11, SQLite FTS5, Rust/napi-rs. Apple Silicon only (M1+, 16GB RAM minimum).

  The main technical challenges were audio capture (resampling, multi-source mixing, format conversion) and keeping the UI responsive while running two AI models — solved by isolating STT and LLM in
  separate Node.js workers.

  MIT licensed. DMG available for download, no code signing yet so you'll need xattr -cr.  I built a macOS app that transcribes meetings in real-time and generates notes (summary, action items, key points) using AI that runs entirely on your machine. No cloud, no API keys, no
  subscription.

  It captures both your microphone and system audio (Teams, Meet, Zoom) simultaneously using a native Rust module with ScreenCaptureKit, then runs Parakeet TDT (NVIDIA, 640MB) for speech-to-text and
  Mistral 7B for summarization — all local via sherpa-onnx and node-llama-cpp.

  Currently French only — multi-language support is on the roadmap. Pull requests are welcome if you want to help add other languages or contribute in any way.

  Stack: Electron 34, Angular 21, NestJS 11, SQLite FTS5, Rust/napi-rs. Apple Silicon only (M1+, 16GB RAM minimum).

  The main technical challenges were audio capture (resampling, multi-source mixing, format conversion) and keeping the UI responsive while running two AI models — solved by isolating STT and LLM in
  separate Node.js workers.

  MIT licensed. DMG available for download, no code signing yet so you'll need xattr -cr.