Offline
The model lives on your laptop.
Recognition runs on the neural engine of an Apple silicon Mac and, on Windows, on the GPU where available. Your audio has no network path to travel.
The default English model downloads once, about a gigabyte, and is then used for every dictation. Other languages are optional downloads from inside the app.
Because nothing is streamed, latency is set by your processor rather than a round-trip to a data center. Modern laptops keep up as you speak.
Optional features that call a cloud language model, such as some rewriting transforms, are clearly labelled and off by default. Read the privacy policy for the exact list.
Related: Why offline matters Privacy policy Transforms