PRODUCT · REVIEWED
What Aurora does
Aurora records a conversation, writes it down as it is spoken, tells the speakers apart and summarizes it — on the iPhone in your hand. This is the whole capability list, in words, with a label on every line.
The shape of it
Aurora is a speech workspace for iPhone. You press record in the room you are already in — nothing joins your call as a participant, and no recording is uploaded to be transcribed. The words arrive as they are spoken, each voice in its own lane. When you stop, a language model on the device writes the session up, and you can ask it what was said.
The home page runs the interface itself, as a replica, so you can watch that happen. Aurora for iPhone is the same product from the other side: the workflow step by step, what a phone needs, and what the build you can install today cannot do.
Three labels are used across this site, and each means one thing. Available in beta is in the TestFlight build and works. Experimental in beta is in the build and is not dependable yet. In development is not in the build at all. There is no fourth label, and no capability on this site goes without one.
Available in beta
- Recording. Aurora listens through the microphone of the iPhone you are holding, in the room or beside the laptop. No bot joins the meeting, and the audio is written to the phone that heard it.
- Live transcription. Speech becomes text while people are still talking. A word is shown as confirmed only once two consecutive decodes agree on it; the unconfirmed tail stays on screen, dimmed, rather than being hidden — the rules are written up here.
- Speaker separation. Aurora tells the people in the room apart as it listens and gives each one a lane. It is not perfect, and we publish what it scores rather than describing it as solved.
- Speaker names the conversation gives. When someone says who they are, a model on the device reads that out of the transcript and puts the name on the lane; otherwise the lanes stay Speaker 1, Speaker 2. The name is stored with that session, and there is no control for typing one in yourself.
- Summaries and action items. When you stop, a language model on the device writes the title, the summary and the follow-ups it heard. Nothing is sent away to produce them.
- Working offline. Recording, transcription, speaker separation and summaries need no connection once the speech models are installed. The models are fetched once, on first launch.
- Your calendar, read on the device. Aurora can show the next event on your calendar. It reads it through the system's own calendar interface, on the phone, and uses no network to do it.
- Plain-text export. A transcript leaves as plain text, to the app or place you choose. Plain text is the only export format Aurora has.
Experimental in beta
These are in the build. They are not finished, and this is the label that says so.
- Ask Aurora. Ask what was decided and the answer cites the moment in this session's transcript that it came from, or says the evidence is not there instead of guessing. The grounding is the session you are in; the wider evidence architecture behind it is not shipped.
- Languages other than English. They work unevenly, and we have no per-language measurement to publish.
- The second pass over a finished transcript. After a session ends Aurora re-reads it. That pass can revise words you watched confirm, and it can relabel a run of speech from one speaker to another.
- Sessions much longer than an hour. They run; they cost more time on older hardware.
In development
Not in the build. Described here so that nobody has to discover it by looking for it.
- Recognizing a known voice in a later meeting. A name is stored with the session it was found in, not learned as a voiceprint, so the next conversation starts from unnamed lanes again.
- Searching across conversations. Search inside a session exists; search across your library does not.
- Syncing conversations between devices. There is no Aurora account, and your library lives in the app on the device that recorded it.
- The Mac app. A macOS target exists in the app's source and there is no release, so everything on this site describes the iPhone.
Where the line is
The whole speech pipeline runs on the device: capture, transcription, speaker separation, summaries, action items and Ask Aurora. Recordings are not uploaded to be transcribed, because no server exists that would receive them.
We will not tell you that nothing ever leaves your phone. Aurora is local-first, not network-free: the app fetches its models once, checks an entitlement when you open a subscription screen, can sync settings to your own iCloud if you turn that on, and sends a diagnostic report when you ask it to. Two of those are worth knowing without reading further. A count of your remaining free minutes is kept in your own iCloud and is the one path that runs on its own. And a diagnostic report started inside a session opens with the switch that attaches that session's content already on — turn it off and the conversation stays where it is.
The Privacy Notice carries the complete list: every path, what each one sends, to whom, and what it keeps. That table lives there and only there, so there is one copy of it to keep true.
What Aurora does not claim
- That nothing ever leaves your device. It is local-first, and the paths that exist are named on the Privacy Notice.
- That transcription or speaker separation is perfect. The measurements are published, including the corpus where roughly two words in three carried the wrong speaker label.
- That a diagnostic report excludes your content by default. Inside a session, it does not.
- That a speaker named in one conversation is recognized in your next meeting, or that conversations sync between your devices.
- A price. Aurora is free while the TestFlight beta lasts, pricing is not decided, and no number on this site is one.
Getting it
Aurora is on TestFlight, free while the beta lasts, and the link is public — there is no waitlist and no account with us.
Or copy the link: https://testflight.apple.com/join/NpdW2ZEb
What the beta covers, what you need and how to install it is a page of its own, and Aurora for iPhone walks through a session as the app runs it.
See also
- Aurora for iPhone — the workflow, the requirements, the acquisition path and the limits of the build you can install.
- Privacy Notice — every network path in full, and what this website stores if you give it an address.
- Try the free beta — the public TestFlight link and the installation steps.
- Engineering notes — first-party measurements of what the app does and what it costs.
Reviewed 21 September 2026 against the Aurora app source. The build changes every week; if a line here is wrong, it is a bug in this page and we want to hear about it at hello@auroraisyours.com.