Lithos Talk is a live conversation translator for iPhone. People you talk with can join from a browser, with no app.
For language pairs downloaded to your iPhone, translation runs entirely on the phone with Apple's on-device speech recognition and translation, and no audio leaves the device (you can require this with the On-device only setting). Apple may collect usage metrics for its on-device translation, such as the language pair and our app's ID, but never what was said or translated. Otherwise, while a translation is running, audio from your microphone is sent over encrypted connections through our server to Google's Gemini on Google Cloud (Vertex AI) to be transcribed and translated, and the translation comes back the same way. Our server passes the audio along in memory only and never stores or logs it; it records only how long each session lasted, for billing. Google doesn't use it to train or improve its products. How long Google may keep it, for example to check for abuse, is set by Google Cloud's terms; see Vertex AI data governance. Older test versions of the app connected to Google's Gemini API directly, where Google keeps prompts and responses for up to 55 days to detect abuse; see the Gemini API usage policies and Generative AI Prohibited Use Policy.
With the Standard voice option (in testing), speech is transcribed by Google, and each finished sentence's text is sent through our server to be translated and voiced by Google Gemini. That text passes through our server in memory only and is never stored by us. The Standard option uses Google's Gemini API, where Google keeps requests for up to 55 days to detect abuse and doesn't use them for training.
If you add your own Gemini API key (Settings → Advanced), that audio goes to Google under your key and your Google project's terms. Your key is stored only in your iPhone's Keychain and is never sent to us.
Some features send your speech, or the text recognised from it, to other companies to be recognised, translated or voiced. Test engines in the app (TestFlight and testing builds, under Settings → Test engines) use them, and so do church rooms run with Live Translation's “Managed by Lithos” option. Which ones a conversation uses depends on its languages, the engine or room settings, and which services are available at the time. They are:
For each of these, we ask for zero data retention where the company offers it (for example, OpenRouter requests for gpt-6-sol and Gemma 4 ask for hosts that state they don't retain data or train on it, and we are asking Google for zero retention on our Google Cloud project). That is requested, not promised: each company's own terms decide what it keeps, and some keep requests for a limited time to detect abuse.
What we store: nothing of what is said, neither audio nor text. Our server passes speech and text along in memory only and never writes them to storage or logs. For a managed church room, the room's audio is relayed through our server to Microsoft (in memory, never stored), so that the church's device holds no Microsoft credential. We keep metadata and counters: how long rooms and connections ran, how many characters, tokens and seconds of audio each service handled per room and language (for the church's invoice and our usage limits), which service answered, and timings. Room names are kept only with those counters, for the church's invoice.
There is no sign-up, email or password. The app creates a random account ID and secret and keeps them in your iCloud Keychain, so a new iPhone on the same Apple ID continues the same account. Our server stores the account ID, a hash of the secret, a short support code, your plan, and how many translation minutes were used each day and month. It also keeps a random device key created by Apple's App Attest for each phone, and the phone's name if iOS provides it.
Subscriptions and minute packs are processed by Apple. We receive Apple's signed record of the purchase (product, dates, renewal and refund status) tagged with your account ID, and never see your name, Apple ID or payment details. To limit repeated free allowances, the app may use Apple's DeviceCheck, which stores one bit per iPhone with Apple, not any identity.
To find and fix problems, the app sends: crash and performance reports from Apple's MetricKit (stack traces, device model, iOS version), and, for each translation session, counts and timings such as the number of turns, reconnects and how long translations took. These are stored without any account or device ID and deleted after 30 days (session counts after 90 days). If you tap Send diagnostics in Settings, the app sends a technical log of the last 15 minutes with your support code; it never includes what anyone said or the translations. None of this contains audio, transcripts or translations.
If you join from a link or QR code, your audio goes to the host's phone, which translates it. The connection between the two devices is end-to-end encrypted: directly between them when possible, otherwise through Cloudflare's relay or our server, which only forward encrypted data. Translation itself is not end-to-end: to translate, the host's phone sends speech to Google Gemini (see above), unless both languages are translated on the host's iPhone. You won't see the conversation as text unless the host chooses to share its transcript; if they do, it's end-to-end encrypted, shown only during the call and never stored. You can add a name so the host knows who's joining; it's optional, sent end-to-end encrypted to the host's phone only, and remembered in your browser or on your phone for next time. The join code expires after the conversation.
Conversations appear on screen while the app is open and are not saved. Your language settings and recent calls (join codes only) are stored on the phone.
No ads, third-party analytics or tracking, and no data is sold or shared for marketing.
Email hello@lithos.community with your support code (Settings → Plan), and your account, device records, usage history and diagnostics will be deleted.