How It Works
The engine behind Millie
Millie turns speech into text using Parakeet v3, an automatic speech recognition (ASR) model, run through FluidAudio's on-device inference engine for macOS. There's no cloud API, no server round-trip, and no account — transcription happens entirely on your Mac.
The speech model
Parakeet v3 is a modern transducer-based ASR model built for accurate, low-latency transcription. Millie runs it locally through FluidAudio, which compiles the model to run on Apple's Neural Engine — the same dedicated silicon that powers on-device machine learning elsewhere in macOS. That's what keeps transcription fast without spinning up your CPU or GPU, and without needing an internet connection once the model is installed.
One-time model download
The first time you run Millie, it downloads the Parakeet v3 model (around 600 MB) from Hugging Face and compiles it for your Mac. This happens once. After that, loading is local and near-instant, and Millie works fully offline.
What happens when you dictate
- Millie captures audio from your microphone in short chunks.
- Each chunk is decoded on-device by Parakeet v3 through FluidAudio.
- The resulting text is copied straight to your clipboard.
Your audio and the resulting transcript never leave your Mac — see the Privacy Policy for details.