Sona for iOS is coming·Join the waitlist →

Notes from SonaVoice

Product updates, engineering deep-dives, and opinions about talking to computers.

Introducing streaming mode

Text now lands as you speak — grammar, punctuation and corrections applied in the stream.

Detecting intent vs. dictation in a live stream, without false positives.

Why a hold-to-talk shortcut beats a wake word, every single time.

Teaching the model that “kubectl” is one token and “dash dash force” is a flag.

Same shortcut, same engine — now on the other half of your team’s laptops.

A dictation tool hears everything. Our architecture makes that boring.

Word error rate hides the errors that matter. Here’s what we measure instead.