Zeza

You talk to a living orb on a single full-screen canvas and watch the words appear as you speak. Then you polish the text, summarise it or change its tone, and send it by message, email or clipboard.

Zeza can also read text back in a natural voice, with the current word highlighted. Every note lands in a local library on your phone, searchable and tagged automatically. There is no account and no backend. Zeza is an iPhone app.

1. Speak. A live transcript appears as you talk.
2. Refine. A second, more accurate pass cleans up the transcript.
3. Shape. Polish it, summarise it or restyle it with a tone preset or your own style.
4. Send or listen. Share it to iMessage, SMS, email or clipboard, or have it read back.

The whole pipeline runs on the phone

Speech, language and voice all run on the device. Live transcription uses the iOS SpeechAnalyzer. The second pass uses the Parakeet TDT 0.6b v3 speech model through FluidAudio. Polishing, summarising, restyling, tagging and spoken follow-up commands run on Apple Foundation Models.

Read-back uses on-device neural text-to-speech: Kokoro-82M for English and Supertonic for Russian, Bulgarian and about 28 more languages, with an Apple voice as fallback. The large models download at runtime and are not bundled in the app. Storage is local SwiftData, so no data leaves the device.

Device requirements

Zeza needs iOS 26 or later on an iPhone that supports Apple Intelligence, with the models fully downloaded.

Zeza is for people who think faster than they type: professionals writing messages and emails on the move, and anyone who wants to capture an idea and send it cleanly. Because everything runs on the phone, it also suits people who handle private or sensitive text.