On-device vs cloud: where your voice actually goes

There’s a lot of hand-waving in voice products about where your audio goes. Here’s the straight version for dolfinspeak.

Local by default

Core transcription runs on your Mac. Your speech becomes text on-device, which keeps the round-trip fast and your raw audio off the network for the common case.

When the cloud shows up

The LLM-powered modes — the ones that reshape a ramble into an email or a tidy list — do more than transcribe, and that work runs in the cloud. That’s the line: transcription is local, transformation is where a server gets involved.

Why the split

It’s a trade between latency, capability, and what has to leave your machine. Drawing the line at “transcribe local, transform remote” keeps the everyday path quick while still letting the heavier features exist.

We’ll keep this post honest as the architecture evolves.