Muesli-HQ/muesli

What it solves

Muesli is a native macOS application designed to provide privacy-focused, local-by-default speech-to-text capabilities. It solves the problem of needing a seamless way to dictate text into any application and transcribe meetings with speaker identification, without relying on cloud services by default.

How it works

The app leverages Apple Silicon's Neural Engine, CoreML, and Metal to run various ASR (Automatic Speech Recognition) models locally on-device. It captures audio from the microphone and system audio (for meetings) using CoreAudio and ScreenCaptureKit. For meeting transcription, it uses VAD (Voice Activity Detection) to split audio at natural speech boundaries and a CoreML-based diarization model to identify different speakers.

Who it’s for

macOS users who require high-performance, private dictation and meeting transcription tools that integrate with their system (e.g., Google Calendar, Apple Shortcuts) and can run entirely offline.

Highlights

• Local-first ASR: Supports a wide array of local models including Parakeet, Nemotron, Whisper, and Gemma 4, with latency as low as 0.13s.
• Meeting Intelligence: Features automatic speaker diarization, camera-based meeting detection, and Google Calendar integration for automatic transcription triggers.
• Join & Transcribe: Extracts meeting URLs from calendar events to launch meetings and start transcription simultaneously.
• Agent-Friendly CLI: Includes a bundled CLI (muesli-cli) that allows coding agents to programmatically transcribe audio files and manage meeting notes via JSON.
• Privacy-Centric: Audio stays on-device by default, with optional BYOK (Bring Your Own Key) integration for OpenAI or OpenRouter for advanced cleanup and summarization.
• System Integration: Deeply integrated with Apple Shortcuts, Siri, and iCloud for text sync between Mac and iPhone.