Sotto records your meetings and lectures, writes them down, works out who said what, and summarises the result — entirely on your iPhone. No account. No server. It works in airplane mode.
Private by architecture, not by policy. Most meeting-notes apps promise not to misuse your recordings. Sotto is built so that there is nothing to misuse: the audio never leaves the phone, so there is no server to breach, no retention schedule to trust, and no policy that could quietly change next year.
You don't have to take that on faith. Turn on Airplane Mode and record a meeting. Everything still works.
Long-form capture built for real meetings and lectures, with live text as you go and recovery from interruptions.
Apple's on-device speech recognition, in around 40 languages, with custom vocabulary for names and jargon.
Neural speaker diarisation on the Neural Engine — the thing Voice Memos and Notes don't do.
Abstract, key points and action items, written by Apple's on-device foundation models. Every action item links back to its moment in the audio.
Text, Markdown, SRT, WebVTT, JSON, and the original audio. All free, all yours, no lock-in.
No sign-up, no email, no analytics, no ads. The app has never heard of you and doesn't need to.
Cloud meeting-notes tools pay for every minute they transcribe, which is why their free tiers are so thin. Sotto's transcription runs on hardware you already own, so each additional hour costs nothing to provide — and charging for it would be an invented constraint rather than a real one.
So recording, transcription, speaker labels and every export format are unlimited and free, permanently. Sotto Pro covers unlimited AI summaries, which are the one genuinely expensive thing the app does, and it pays for the app's continued development. The free tier includes ten summaries a month.
Telling voices apart is the hard part and the part most likely to be oversold, so it is measured rather than asserted. These are scores against the AMI Meeting Corpus and its human annotation, using diarisation error rate — lower is better. Two of the four meetings were held out entirely during tuning.
| Meeting | Recording condition | Speakers found | Error rate |
|---|---|---|---|
| ES2004a | Close-talk microphones | 4 / 4 | 9.7% |
| ES2004a | One distant microphone | 4 / 4 | 16.3% |
| IS1009a (held out) | One distant microphone | 4 / 4 | 24.2% |
| TS3003a (held out) | One distant microphone | 4 / 4 | 20.7% |
An earlier fallback that guessed at speakers without the neural models scored close to a coin flip, so it was removed rather than tuned. When Sotto can't identify speakers, it says so instead of guessing — for a therapist or a lawyer, a confident wrong attribution is worse than none.
Worth knowing before you rely on it: