Privacy
What poly does with what you say.
poly translates speech in real time. This page explains, in plain language, what the app and the servers behind it do with a conversation — and what they do not. The short version is that conversation content is processed in memory, on our own servers in Switzerland, and is never stored.
- Updated
- — October 21, 2026
- Applies to
- — the poly iOS app and this site
In short
- Nothing is recorded.
- Audio, transcripts and translations exist only while a conversation is open. Nothing is written to disk — not on our servers, not in the app, not in a backup.
- Nothing leaves Switzerland.
- Speech recognition, translation and speech synthesis all run on models we host ourselves, on servers we administer here. No third-party translation or speech service is involved.
- Nothing identifies you.
- There is no account and no sign-up. The access token says which institution a session belongs to, not who you are.
- Nothing trains a model.
- Conversation content is never used to train, tune or evaluate our models.
Who is responsible
poly is an independent Swiss AI research lab in Zürich, run by Jonathan Gosteli, who is responsible for the processing described here. Questions about privacy, and requests about your data, go to:
What this page covers
The poly iOS app, which staff at a small number of partner institutions install from the App Store; the backend the app connects to; and this website. The app is unlisted: no search or chart in the App Store shows it, it is reached by a link given to each institution, and it is not offered to the public.
What the app processes
While a turn is being spoken, four things:
- Microphone audio, captured only while you are recording, streamed over an encrypted connection to our servers, and processed as it arrives.
- The transcript and the translation produced from that audio — shown on screen so both speakers can check them, and spoken aloud.
- The two languages of the conversation, which are chosen before it starts.
- An access token, issued to the institution that invited you and kept in your device's keychain on that device alone. It records which institution a session belongs to. It does not record who you are.
The app has no account, no sign-up and no profile. It never asks for a name, an email address, a date of birth, or any patient, client or case identifier — there is no field in which to enter one.
What we use it for
One purpose: turning speech into translated speech, live. Audio goes to speech recognition, the resulting text to machine translation, the translated text to speech synthesis, and the spoken result comes back to the app. That is the entire chain.
We do not use conversation content to train, tune or evaluate models. We do not profile you, advertise to you, sell anything about you, or make automated decisions about you. There is no tracking in the app.
How long we keep it
We don't. On our servers, audio, transcripts and translations live in memory while a turn is being translated and are discarded when it is done. In the app, the conversation on screen — and the translated speech kept so that a turn can be played again — stays in the app's memory until a new conversation is started or the app is closed. None of it is ever written to disk — not on our servers, not in the app, not in a backup or an archive.
That has a consequence worth stating plainly: we cannot hand a past conversation to you, to your institution, or to anyone else, because no copy of it exists anywhere.
What we do record
Operational records that carry no conversation content: when a turn happened, how long each stage took, the language pair, error codes, and the label of the institution the session belongs to. They tell us whether the service is healthy and fast, and how much each institution uses it. They are kept for up to 400 days, and backed up weekly, encrypted, to a computer the lab administers.
Audio, transcripts and translations never appear in them, and we treat that as something to verify rather than assume: an automated check runs a real turn through the system and then searches the servers' logs and disks for the words it produced.
Where it is processed
On servers we administer in Switzerland. All three stages run there, on open-weight models we host ourselves. Conversation content is never sent to a third-party translation or speech service and is not transferred abroad. The app does not use Apple's or Google's dictation either.
Who else is involved
Apple distributes the app through the App Store, and test builds through TestFlight. Apple processes your Apple Account's download of the app, the email address a TestFlight invitation was sent to, and, if you allow it on your device, crash and usage data about the app; that processing is Apple's own and is covered by Apple's privacy policy. Apple never receives conversation content.
Vercel hosts this website and processes what any web server sees — IP address, browser, page requested — in order to serve the page. It also passes the website's request-access form on to us, from a server in Frankfurt, without keeping it. It has no part in the app.
Infomaniak, in Switzerland, hosts our email: what you write to us, and requests sent with the website's form. It has no part in the app either.
No one else. The app contains no third-party code at all: no analytics, no advertising, no crash reporting of our own.
Sensitive conversations
Conversations in care, legal and administrative settings can contain health data and other sensitive personal data. That is exactly why the system is built to keep nothing: the strongest protection a conversation can have is that no copy of it outlives it.
A communication aid, not a medical device
poly is a communication aid. It is not a medical device, it must not be the basis for a clinical decision, and it does not replace a qualified interpreter. The transcript on screen exists so that both speakers can see what the system heard and catch an error before it matters.
Your rights
Swiss data protection law gives you the right to ask what we process about you, to have inaccurate data corrected, to have data deleted, and to object to processing. Write to us and we will answer.
Two honest notes. Because conversation content is not retained, a request about it can only be answered with "there is nothing to give you" — there is no store to search. And if you use poly through your employer or another institution, that institution chose to use it and is your first point of contact; a written agreement sets out what each of us is responsible for.
Security
Everything between the app and our servers travels over an encrypted connection. Access is by per-institution token, so one institution's token can be withdrawn without affecting anyone else. The servers are ours, in Switzerland, and access to them stays inside the lab.
Children
poly is provided to professionals at partner institutions for their work. It is not directed at children and not offered to the general public.
This website
No cookies, and no tracking across sites. Page views are counted with Vercel Analytics, which works without cookies and builds no profile of you.
The "Request access" form sends us the email address you enter, and nothing else. It reaches us as an email in our mailbox and is not stored anywhere on the way. We use it only to answer your request, and delete it once we no longer need it for that, or sooner if you ask.
Changes
When the system changes, this page changes with it and the date at the top moves. Pilot institutions hear about material changes directly.
Contact
Anything on this page, and anything it does not answer: