Privacy Policy

1. Introduction

Unitra, Inc. ("Unitra," "we," "us," or "our") operates the Unitra desktop application and cloud services that provide real-time voice translation and translation of text shown on your screen, which the app recognises on your device. This Privacy Policy explains how we collect, use, share, and protect your information when you use our services.

Our legal grounds for each kind of processing are summarised in Section 2. Where the law requires your consent (most importantly, before we can keep your translation text, audio or screen samples for model improvement), we ask for an explicit, affirmative opt-in. Continuing to use the service is not treated as consent for optional processing. You can review, change, or withdraw your choices at any time in your account settings.

2. Why we are allowed to process your data

Most data protection laws require us to identify a legal ground for each kind of processing. In plain terms:

• To deliver what you signed up for. We process the information needed to translate your requests, manage your account, enforce subscription limits, and bill you. This includes, when your PC is under heavy load, running one step of speech recognition on our own private cloud GPU servers from speech features (see Cloud assist in Section 3; you can switch it off). Without this, we couldn't provide the service.
• To keep the service running and safe. We process a small amount of operational information for short-retention server logs, security and abuse prevention, rate-limit enforcement, debugging, checking that app updates install correctly, and aggregate service-health metrics. The data is minimised, kept only as long as needed for the purpose, and access is restricted. You can object to this kind of processing at any time.
• With your consent, to improve our models. We collect translation text, audio samples and screen samples for model improvement only when you turn on the single Help Improve Unitra opt-in (Settings › Privacy). It is off by default. You can withdraw it any time; new collection stops as soon as our systems propagate the change, typically within minutes. A request that was already in flight when you withdrew may finish and be stored, but no further samples will be taken after that. (Withdrawing doesn't make our prior collection unlawful, but you can request deletion of what we already collected; see Section 8.) While we offer it, turning this switch on also doubles the daily translation allotment on plans that have a daily limit. Your plan’s own allotment never depends on consenting, and withdrawing only removes the bonus going forward.
• To understand which channels bring people to Unitra. On the website we rely on our legitimate interest in measuring our own marketing (Section 5): a random identifier in your browser, how you reached the site, and whether a download led to an installation and an account. We do not run this measurement at all for visitors in the EEA, the United Kingdom or Switzerland, where storing such an identifier would need your consent, and elsewhere you can object at any time with the opt-out in Section 5. Reporting an ad click's outcome back to the platform that served the ad rests on the same interest, is limited to visitors outside those regions, and stops when your browser sends the Global Privacy Control signal.
• To meet our legal obligations. For example, retaining payment records for tax purposes or responding to lawful requests.
• To send you news and offers, where the law where you live allows it. This is the only processing whose ground depends on your country. In the EEA, the UK, Switzerland, Korea, Japan, China, Russia, Canada, Australia and New Zealand we rely on your consent, and the box is unticked until you tick it. In the United States and a small number of other countries whose law permits commercial email with a working unsubscribe, we rely on legitimate interest and the box starts ticked, which you can untick during sign-up or at any time afterwards. Either way it is one click to stop, and stopping never affects account, security, or billing notices. See Section 4a.

Why translation text, audio and screen samples share a single switch

The text of a translation request and the audio behind it describe the same activity: what you said, and what we returned. They are useful for model improvement only together: a transcript without the matching audio teaches our speech recogniser nothing, and audio without the matching text teaches our translator nothing. We therefore treat them as one processing operation governed by one switch. If you want to allow text contributions but not audio (or vice versa) for your specific account, email info@unitra.ai and we will configure it for you.

Screen samples do the same for screen translation: a sample is the picture the recognised text was read from, so we can tell whether a mistake came from reading the screen or from translating it. Because screen samples widen what the switch covers, people who turned it on earlier are asked to confirm it again when they update to a version of the app that takes screen samples (Section 13). If you were asked to confirm again and choose to decide later, your text and audio contributions continue and no screen samples are taken.

Where consent is the legal ground for keeping data, it is the only ground we use; we do not fall back on "legitimate interest" to keep collecting from people who have not opted in or who have opted out.

3. What we collect

What's on by default, at a glance

Account information

When you create an account we collect your email address and display name. Your password is stored using one-way encryption that is intentionally slow to compute, so even in the worst case where our database were compromised, your original password could not be recovered. If you sign in with Google or Steam instead, we receive from that provider only what we need to create and link your account: from Google, your email address and name; from Steam, your Steam ID and profile name. See Section 6.

Translation data

When you use our cloud translation service, the text content of your translation request is transmitted to our servers and forwarded to the translation model providers listed in Section 6, solely to return a translation. By default, without your explicit opt-in, translation text is processed in real time and is not written to long-term storage. Brief copies may exist for a few seconds (occasionally up to a few minutes) in the components that move the request through the system, and are discarded as soon as the request completes.

If you turn on Help Improve Unitra in Settings › Privacy, translation text you submit after that point is stored in our database under a random reference number rather than your email or name (this is what data-protection law calls "pseudonymisation"; see Section 7). You can review, export, or delete it at any time. Each stored sample also includes a small amount of operational information about how the request was processed (for example, recognition confidence, request timing, and your subscription tier at the time), used solely to evaluate and improve model quality.

Audio

By default, Unitra recognises speech entirely on your device; your audio is not transmitted to our servers (the one assisted step described under Cloud assist sends speech features). If you turn on Help Improve Unitra (the same single switch described above), short clips of your own microphone audio captured during normal use may be uploaded solely to evaluate and improve recognition accuracy. They stay linked to your account for up to 90 days (you can delete them at any time during that period), after which they are de-identified and retained for at most 2 years before permanent deletion (see Section 7). We never upload audio of other people on your device (for example, teammates on a call). Audio is not used for biometric identification or speaker recognition (see Section 8 for the Illinois-specific commitments).

Cloud assist

Speech recognition in the desktop app runs in two steps: the first turns a few seconds of audio into speech features (a numeric summary of the sound, computed on your device), the second turns those features into text. The second step needs a capable graphics processor. When your PC's graphics processor is busy with a game or short of memory, or has no suitable graphics processor at all, the app may hand that second step to one of our own private cloud GPU servers, dedicated machines operated solely by Unitra, so captions keep up. It is used only while that is the case and switches back to on-device recognition as soon as your PC recovers.

What is sent. Only the speech features for the current utterance, over an encrypted connection (TLS) through our edge provider Cloudflare to one of our private GPU servers in the region nearest you (Asia, North America or Europe), together with the languages you have selected. These servers run on dedicated machines we rent from Google Cloud or Amazon Web Services; the provider supplies the hardware and network, and only Unitra operates the server. Speech features can be turned back into what was said, so we protect them as your voice data: they are processed in our server's memory only, discarded the moment the text is returned, and are not stored, logged, used to improve our models, or shared with anyone. The returned text is then handled exactly like text recognised on your device (and, if you have Help Improve Unitra on, under the same rules as other translation text).

Control. Cloud assist is on by default. The first time it is used on a device the app shows a notice explaining this. You can turn it off at any time in Settings › Privacy › Cloud assist; recognition then stays on your device even when that makes captions slower. The app also records, as part of the session statistics described below, whether and for how long cloud assist was used in a session (no content).

Not a voice identifier. Speech features describe the words spoken, not who spoke them, and we do not use them for speaker recognition or any biometric purpose (Section 8).

Screen text recognition

The desktop app can also translate text shown on your screen, such as in-game chat. When you use this, the app captures only the area of the screen you select (the one exception is the small game window picture described under Screen samples), recognises the text in it on your device, and sends only the recognised text to our servers as translation text; the Translation data rules above apply to it in the same way. The screen images themselves are processed on your device and do not leave it, except as the screen samples described below.

If Help Improve Unitra is on, the recognised text and its translation are kept as ordinary translation text samples under the same rules as other translation text (Section 7). Player names in chat lines are removed first.

Screen samples (Help Improve Unitra)

If Help Improve Unitra is on, the desktop app may occasionally keep a screen sample from a screen translation session and upload it to us. A screen sample is either an image of the area you selected, together with the text recognised in it and where each line was, or a small, low-resolution picture of the game window (at most a quarter of the window's resolution, with the area you selected blacked out), which shows where the game places its chat and subtitles. Each sample carries the same technical details as the session statistics described below.

Why, and how many. We use screen samples only to improve how Unitra recognises text on screen and to work out better screen areas and image settings for each game. The app keeps only a small number per session, with a daily limit, chosen mostly from moments when recognition struggled. Samples are taken only while a game is the window in front on that display, never while another program, including Unitra, is in front of it; any part of your selected area outside the game window is blacked out.

Other people. A sample can show other players. The app blacks out player names in chat before uploading a sample and removes them from the recognised text; other content, such as chat messages, may remain. We do not run face detection or recognition, and we never use screen samples to identify or profile anyone, publish them, or sell them. If you think you appear in someone's screen sample, email info@unitra.ai with the game, your in-game name and the approximate dates, and we will delete the samples we can match.

Storage and control. Samples are encrypted on your device until they are uploaded and stored encrypted with us (Sections 6 and 10). They stay linked to your account for up to 90 days and are deleted no later than 1 year after collection (Section 7). Samples waiting on your device belong to the account that was signed in when they were taken: signing out deletes them, and they are never sent under another account. Turning Help Improve Unitra off stops new samples and deletes those still waiting on your device; the delete control in Settings › Privacy, or deleting your account, removes the uploaded ones.

Usage data

We collect operational information about requests to our cloud services, including:

• Number of translation requests and character counts
• Source and target language pairs
• Your subscription tier and usage against limits
• Timestamps and timing measurements: how long the request took to answer, and, for the desktop app, how long each step of a translation took (for example: the pause before your sentence was treated as finished, speech recognition on your device, the network round trip, and the moment the result was drawn on screen). These are durations in milliseconds, never the words themselves.
• The country your request originated from (2-letter country code only, provided by our edge provider Cloudflare, e.g. "JP", "US") and the identifier of the Cloudflare data centre that handled the request (3-letter code, e.g. "NRT" for Tokyo, "SJC" for San Jose). This is country-level coarse only; we do not collect city, latitude/longitude, or the raw IP address that resolved to this country.

This is used to operate, secure, and optimise the service, and to enforce subscription limits. It is not used to train translation models.

The country and data-centre fields above are used so we can monitor whether the service is healthy in each region (request volume, latency, error rate per region) and route traffic accordingly. They are kept linked to your account for 90 days, after which they are automatically cleared from individual request records; only aggregate per-region statistics survive past that point. We do not use them to build a profile of your movements or to personalise the product.

Session statistics from the desktop app

Speech recognition runs on your device, so our servers cannot see whether a session was slow or whether there was anything to translate at all. To answer those questions, the desktop application sends us one small statistics record per session when the session ends. It contains: a random identifier for that one session (so a record that reaches us twice is not counted twice), when the session started and how long it lasted, how many seconds of speech were heard on each side (your microphone, and the other people), how many translations were produced in each direction, whether the two directions ever produced a translation close together, which game profile was active, which translation engine served the session (on your device, in the cloud, or both), the app version, and, for each step of the pipeline, how many utterances there were and how long that step typically took (its median and 95th-percentile duration for that session). That list is the whole record.

When you use screen translation, the app sends one similar record per screen translation session. It contains: a random identifier for the session and for the run of sessions it belongs to (choosing a new area starts a new session in the same run), whether it resumed after choosing a new area, when it started, how long it lasted and how it ended, and the app version; the screen translation settings and recognition engine, and how many languages you selected (never which ones); how the screen was captured; your display's resolution, scaling, rotation and layout, and whether HDR is on, its refresh rate and colour format; the position and size of the selected area, as fractions of the screen and of the game window; the file name of the program in front on that display (for example overwatch.exe), our name for the game if we support it, its window mode and size, and for how much of the session it was in front; and recognition counts and timings, including summary counts of recognition confidence, text size and contrast and a coarse grid of where on the screen text appeared. That list is the whole record. It never includes window titles, folder paths, or a list of the other programs you are running.

Neither record ever contains what was said, recognised on screen or translated, any audio or screen image, which languages were spoken or recognised, the names of other players, or a timestamp for individual sentences: the timings are computed on your device and only the per-session summary leaves it. We use these records solely to see whether Unitra helps in real situations and to find out where the pipeline is slow, on the same legitimate-interest basis as the crash reports below. They stay linked to your account for up to 90 days. After that we permanently remove the link to your account and the session identifiers, round times to the day, durations to the minute and screen positions to a coarse grid, and keep the records only in that de-identified form. Sharing session statistics is on by default and you can turn it off at any time in Settings › Privacy (“Share session statistics”); the change takes effect immediately, and the record for a session in progress is discarded.

Device information

We collect basic device information sent with each request, such as your operating system and version (e.g., "Windows 11"), the application version, the type of processor (e.g., "x86" or "Apple silicon"), and a general description of your graphics processor (e.g., "NVIDIA RTX 4070"), when available. We use this to debug platform-specific issues and to understand how the service performs on different hardware.

We do not combine these fields, including the display details, into a stable device identifier; we do not use third-party analytics, tracking pixels, or browser/device fingerprinting. The only identifiers we assign ourselves are the random ones described under Crash reports below and in Section 5. If you have opted in to Help Improve Unitra, a copy of these fields is kept next to each translation sample so we can group quality metrics by device after the sample is no longer linked to your account.

Crash reports

The Unitra desktop application can send a crash report to our servers when something goes wrong. A crash report contains:

• Technical details of the error (such as the error message and where in the application it happened);
• A short list of the actions you took just before the crash;
• A random number that identifies the installation (so we can tell two crashes came from the same copy of the app, without learning who you are);
• Basic device and version information; and
• For the most severe crashes, a small diagnostic snapshot of the application's state at the time it crashed.

Before sending, the crash reporter strips known parts of memory that may hold translation text, audio, or login credentials; we discard any such fields server-side as well. We do not use crash reports to recover translation content.

The same Settings › Privacy switch also controls a small installation report: on its first launch the app tells us the random installation number above, the download id from the installer's file name if there is one (see Section 5), the app version, operating system version and system locale, and later whether signing in succeeded and by which method (email, Google or Steam). It carries no translation content and no hardware details.

When it's on. Crash reporting is on by default so we can detect and fix problems across the user base. Without it, the only crashes we'd ever see are the ones users take the time to report manually, and most never do. You can turn it off at any time in Settings › Privacy; the change takes effect the next time you start the application.

Why on by default. Crash reports are minimal (no translation text, no audio, no account-identifying data), kept for a short period, and used solely to improve reliability. This is the same legitimate-interest basis that any responsibly-built productivity application relies on for stability telemetry. If that doesn't sit well with you, the off switch is one click away.

How long we keep it. Up to 90 days. After that we strip the installation identifier and only keep aggregated counts grouped by the type of crash. Crash reports are used solely to fix bugs and measure how reliable each release is, never for marketing and never for training our models.

Update diagnostics

When the desktop application checks for, downloads, or installs an update, it sends us a small report of how that went: whether the step succeeded or failed, and if it failed, where and why. A report contains: the version you were on and the version being installed, your operating system and its version, the processor type, the release channel (for example, stable or beta), whether the app came from our website or from Steam, which step of the update it describes (checking, downloading, installing, or confirming that the new version started), a short error type and error message, how many attempts it took and how long the step lasted, and the same hashed installation identifier that crash reports use. That list is the whole report.

Before a report leaves your device, the application removes your user name and personal folder paths from the error message (for example, C:\Users\YourName becomes %USERPROFILE%) and shortens it to 500 characters; our servers remove them again on arrival. A report never contains your account, your email address, what you translated, or any audio, and we do not store your IP address with it. We use these reports solely to find out whether an update reached people and to fix the ones that did not, on the same legitimate-interest basis as crash reports.

The same switch as crash reports. Update diagnostics are on by default and are controlled by the crash reports switch in Settings › Privacy. Turning it off stops these reports and discards any that are still waiting on your device to be sent. We keep them for up to 90 days.

Update checks

Copies of Unitra installed from our website check updates.unitra.ai for new versions (copies installed through Steam are updated by Steam instead). Each check sends the app version, the release channel, your operating system and processor type, and the hashed installation identifier. We release new versions gradually, and the identifier is what lets our update server decide whether your copy is offered a new version yet, so the same copy gets the same answer each time. The update server uses it only to make that decision and does not store it. Because the app cannot keep itself up to date without these checks, they are not tied to a privacy switch.

When you uninstall the app

On Windows, when you remove Unitra with its uninstaller, the uninstaller sends us one short notice that the app was removed. It contains the hashed installation identifier, the app version, the release channel, whether the app came from our website or from Steam, and whether the uninstaller was run without on-screen prompts. It contains nothing about your account or how you used the app. It is not sent when an update replaces the app with a newer version. We use it to understand how many people stop using Unitra after each release, on the same legitimate-interest basis as the session statistics above.

The notice is controlled by the session statistics switch in Settings › Privacy (“Share session statistics”). While it is off, the application does not prepare the notice, so uninstalling sends nothing. We keep these notices for up to 90 days.

Uninstall survey

When you uninstall the app yourself, the uninstaller may open a short survey on unitra.ai in your browser asking why you are leaving. Answering is entirely optional: you can close the page and nothing is sent. If you do answer, we receive the reason you picked, the comment you wrote (if any, up to 500 characters), the language of the page, the app version, and the hashed installation identifier when the uninstaller passed it along, which lets us connect your answer to the uninstall notice above. Please do not put personal information such as your name or email address in the comment. We use the answers solely to improve Unitra and keep them for up to 12 months (see Section 7).

Session quality feedback

After some sessions the desktop app may ask how the session went. Answering is optional. If you answer, we receive your rating, reasons and comment, together with technical details of that session of the same kind as the session statistics above (for screen translation, also the game program and the selected area). The words that were said or recognised are included only if you have turned on Help Improve Unitra. We use feedback solely to improve Unitra and keep it for up to 3 years (see Section 7).

Payment information

We never receive, store, or have access to your credit-card numbers or bank-account details. How a payment is handled depends on where you buy:

• Checkout on the web is processed by Stripe. Purchases may be sold through Link, a Stripe service acting as reseller and merchant of record; Link then collects your payment and billing details, calculates tax, sends your receipts, and handles payment support as an independent controller under its own privacy policy. We receive your subscription or purchase status, the plan, price, currency and billing country, and the Stripe identifiers needed to connect the payment to your account and to process refunds.
• Purchases in the Steam version of the application are sold by Valve Corporation, which processes the payment as an independent controller under its own privacy policy. We receive your Steam ID, the order and agreement identifiers, the item, price, currency, country, and the status of the transaction, including later renewals, refunds, and chargebacks.

We keep these payment records for as long as tax and accounting law requires (see Section 7).

4. How we use it

We use the information we collect to:

• Provide the service: process your translation requests, manage your account, deliver subscription features.
• Enforce usage limits: track your usage against your subscription tier.
• Operate and protect the service: rate limiting, abuse prevention, fraud detection, debugging, and aggregate service-health analytics.
• Improve our translation and speech-recognition models: only using data from people who have explicitly turned on Help Improve Unitra.
• Communicate with you: essential account notifications such as password resets, security alerts, and material policy changes.
• Tell you about Unitra: occasional news, offers, and updates about new features, sent separately from the notices above and only while news and offers are switched on for your account. See Section 4a.
• Measure our own marketing: learn which websites, campaigns and ads bring people who install Unitra and create an account, and report an ad click's outcome to the platform that served it (outside the EEA, UK and Switzerland). See Section 5.
• Comply with law: respond to lawful requests, enforce our Terms, and meet tax / accounting obligations.

4a. News and offers

This is the only email we send that is not strictly about running your account: occasional news, promotions, offers, and updates about new features. It goes out rarely. Two limits hold whatever we send under this heading, and both are repeated in every message: it is always about Unitra itself, never a third party's advertising, and we never rent, sell, or lend your address to anyone else for their own marketing.

Whether you start switched on

That depends on where you are when you create the account, because the rule differs by country:

• Off until you ask for it: the EEA, the United Kingdom, Switzerland, Korea, Japan, China, Russia, Canada, Australia, New Zealand, and every country we have not specifically classified. The sign-up box is presented unticked. We will not send you anything under this heading unless you tick it.
• On unless you say otherwise: the United States and its territories, Singapore, Hong Kong, Taiwan, Brazil, Mexico, and Argentina, whose law permits commercial email with a working unsubscribe. The sign-up box is presented already ticked, and you can untick it there or unsubscribe from any message.

If we cannot determine your country, we treat you as being in the first group. The country that decides this is recorded once, when the account is created or when you change the setting. It is not re-evaluated later, so travelling does not change how we treat your account.

How to stop

Every message of this kind carries a one-click unsubscribe link, the standard one-click unsubscribe header that mail applications use, and our name and postal address. You can also switch them off at any time on your account page at auth.unitra.ai/profile. We action an unsubscribe immediately. Account, security, and billing notices are a separate category and keep arriving.

Records we keep

We record when your setting changed, whether it was you or the regional default that set it, the country whose rule applied, and the version of this policy in force at the time. We keep these records for the same period as our other consent records (see Section 7), because in several countries the burden of proving that an address agreed to receive commercial email falls on the sender.

Where the law where you live sets a period after which that agreement has to be renewed, we record when yours falls due and ask you to renew it. In Korea, for example, that period is two years: once it passes we show you a reminder on your account page at auth.unitra.ai/profile asking you to confirm again. Confirming restarts the two years. You can also unsubscribe at any time, from that page or from the one-click link in any message.

5. Cookies and local storage

Our website (unitra.ai) does not set any first-party cookies and does not use third-party tracking scripts. We do not use analytics services such as Google Analytics. We do run a small measurement of our own, described under Site measurement below.

Our infrastructure provider, Cloudflare, may set a small number of strictly-necessary cookies for security purposes such as bot detection and protection against denial-of-service attacks. These cookies are essential for the service to operate and cannot be disabled.

Site measurement (first-party)

Not in the EEA, the United Kingdom or Switzerland. Visitors from those regions are recognised at our edge network from the country of the request and the measurement script stays switched off for them: nothing is stored in the browser and nothing is sent. The country is used for that one decision and is not kept.

To learn which channels bring people to Unitra, the website stores a random identifier in your browser's local storage (not a cookie), together with how you reached the site: the page or advertising link you came from and the page you landed on. Each record also notes the type of device the page was opened on (Windows, Mac, Linux, ChromeOS, Android, iOS or other), taken from the description your browser sends with every page; nothing more specific, such as the model or the browser version, is kept. We use it to tell visits from phones, which cannot run the desktop app, apart from computers. The identifier is kept for up to 90 days after your last visit. When you download the installer we generate a random download number and put it in the installer's file name. The installer reports with that number when it is opened, whether the installation finished, and if it did not, the likely reason (for example too little disk space, or the app still running) together with the free space on the installation drive; on its first launch the app tells us which visit led to that installation and afterwards whether sign-in succeeded.

We use this only to measure our own marketing. We do not store your IP address, do not fingerprint your browser, and do not use third-party analytics. If you arrived through an advertisement and you are outside the EEA, the United Kingdom and Switzerland, we may report back to that advertising platform (for example Google Ads or Reddit) that its click led to an installation, an account, a first completed translation or a purchase, using the click identifier the platform itself placed in the link. We report only for the countries we have switched this on for (at present the United States); everywhere else, and always in the EEA, the United Kingdom and Switzerland, the click identifier stays with us. We never send your email address, name, account details or anything you translate, and we do not report anything for browsers that send the Global Privacy Control signal. You can opt out of all of this at any time with the button below; the site and the download work the same either way. Retention is described in Section 7.

Opt out of site measurement

The Unitra desktop application does not set or read website cookies. It does store a login token locally on your device, the same way websites store sign-in cookies, so you don't have to log in every time you open the app. You can clear it by signing out of the desktop application or by deleting your account.

Do Not Track and Global Privacy Control (GPC): we do not track you across other sites. The one thing these signals change is the conversion reporting described above: if your browser sends GPC or DNT, we do not report your click back to an advertising platform.

6. Who we share it with

We share data with a limited set of third-party service providers ("sub-processors") solely to provide our core functionality. Each one is bound by a written data-processing agreement that restricts their use of your data to the purposes we specify and requires appropriate security and confidentiality measures.

Our current sub-processors are:

• Amazon Web Services, Inc. (United States; GPU servers also in Asia): one of our translation providers (Amazon Translate); translation text is sent to it solely to return a translation and is not used to improve or train AWS services. Some of our private GPU servers for cloud assist run on dedicated machines rented from it; AWS supplies the hardware and network only, the servers are operated solely by Unitra, and speech features are processed in memory only.
• Cloudflare, Inc. (United States): edge network, security and bot protection, hosting for our websites, admin console and parts of our backend (including the databases for billing records and for our creator program), storage (R2) for opted-in audio samples, opted-in screen sample images and crash diagnostics, and transactional email delivery for account verification, password resets, security alerts, and billing notifications. Cloudflare is also one of our translation providers: translation text may be sent to its Workers AI inference service solely to return a translation. Connections for cloud assist are routed through Cloudflare's network to our private GPU servers; Cloudflare relays the encrypted speech features and does not store them.
• DigitalOcean, LLC (United States): application hosting and managed databases. This is where account information, usage records, opted-in translation samples (including the text recognised in opted-in screen samples), and consent records are stored. Its AI inference service is also used to review publicly available creator profiles for our creator program, and is a standby translation provider that receives translation text only if our primary providers are unavailable.
• Google Cloud (United States, European Union and Asia): runs the language models (Gemini, through Vertex AI) that translate most requests. Requests are routed to the region nearest you, and some use Google's global endpoint. Most of our private GPU servers for cloud assist run on dedicated machines rented from it, in the region nearest you; Google supplies the hardware and network only, the servers are operated solely by Unitra, and speech features are processed in memory only.
• Microsoft Azure (global): one of our translation providers (Azure AI Translator); requests are processed in the Azure region nearest the request.
• Modal Labs, Inc. (United States): cloud inference for our speech-recognition and translation models.
• Resend (Plus Five Five, Inc.) (United States): delivery of news and offers email only, from a dedicated sending subdomain. It never receives translation text, audio, or payment data, and it is not used for account, security, or billing notices, which continue to go through Cloudflare. Only the addresses of people who currently have news and offers switched on are sent to it. (Resend Privacy Policy)
• Stripe, Inc. (United States): subscription billing, payment processing, and tax calculation. (Stripe Privacy Policy)

Sellers of record. When a purchase is sold through Link (Stripe) or through Steam by Valve Corporation, that company is the seller and processes your payment details as an independent controller, not as our sub-processor; its own privacy policy applies to that processing (Link Privacy Policy, Valve Privacy Policy). We exchange with it only the transaction information described in Section 3.

Sign-in providers. If you choose to sign in with Google or Steam (Valve Corporation), that provider confirms your identity to us as an independent controller under its own privacy policy (Google Privacy Policy, Valve Privacy Policy). We receive only the account details described in Section 3 and do not send your Unitra data to the provider.

Advertising platforms (Google Ads, Reddit) are not sub-processors. If you reached unitra.ai through one of their ads and you are outside the EEA, the United Kingdom and Switzerland, we may send that platform its own click identifier together with the fact that the click led to an installation, an account, a first completed translation or a purchase, so it can measure the campaign. At present this is switched on only for visitors in the United States. It receives no email address, name, account identifier, usage data or translation content from us, and browsers that send the Global Privacy Control signal are excluded. See Section 5.

A current, itemised list of sub-processors, including each entity's name, purpose, and processing location, is maintained at unitra.ai/sub-processors. We update that list whenever we add, remove, or replace a sub-processor. If you would like to be told by email when it changes, or want to object to a sub-processor, contact us as described in Section 14.

We do not sell or rent your personal data, and we share nothing with advertisers or data brokers beyond the conversion reports described above. We do not use third-party analytics services; the site measurement described in Section 5 runs on our own servers. We do not share your data with any parties beyond those described above, except where required by law or to protect the rights, safety, or property of Unitra or others.

7. How long we keep things

We keep personal data only for as long as necessary for the purpose it was collected for.

• Translation text, opted-in users (0–90 days): stored encrypted, with the link to your account replaced by a random reference number ("pseudonymisation"). You can view, export, or delete it at any time in Settings › Privacy.
• Translation text, opted-in users (after 90 days): we permanently sever the link between the sample and your account so we no longer know who it came from. The text content itself, however, may still contain incidental personal information that you happen to have included (for example, a name in a sentence you translated). For that reason we treat the post-90-day dataset as pseudonymised, not "anonymous", and continue to apply the same security and access controls. We may retain it for ongoing model training and evaluation. If you don't want any of your samples to reach this stage, delete them within 90 days using the controls in Settings › Privacy.
• Translation text, users who have not opted in: not stored beyond the brief processing needed to return a result (typically seconds; at most a few minutes inside the components moving the request through the system).
• Audio, opted-in users (0–90 days): stored encrypted, with the link to your account held as a random reference number ("pseudonymisation"). You can export or delete it at any time in Settings › Privacy.
• Audio, opted-in users (after 90 days): we permanently sever the link between the recording and your account. Because a voice recording is inherently harder to strip of identifying characteristics than text, we do not claim the result is fully "anonymous"; we treat it as de-identified: the account link and request metadata are removed, and the same encryption and access controls continue to apply. In this de-identified state we retain audio solely to train and evaluate our speech-recognition models, for a maximum of 2 years from the date of collection, after which it is permanently and irreversibly deleted. If you don't want any of your recordings to reach this stage, delete them within the first 90 days.
• Speech features (cloud assist): never stored. They exist only in the memory of our private GPU server for the duration of one recognition step (a few hundred milliseconds) and are discarded when the text is returned; our servers keep no log of their content.
• Screen samples, opted-in users: images and their recognised text are stored encrypted, linked to your account by a random reference number ("pseudonymisation") for up to 90 days, during which you can export or delete them in Settings › Privacy. After that the link is permanently removed; because a sample can show other people's names or messages, we treat the result as de-identified, not "anonymous", keep it solely to improve text recognition, and delete it no later than 1 year after collection.
• Upload records (opted-in audio and screen samples): the IP address an upload came from is stored with it for security and abuse prevention, and is cleared when the link to your account is removed (at most 90 days) or the upload is deleted.
• Usage records: individual request records kept up to 90 days for billing, quota enforcement, and security; afterwards aggregated into anonymous statistics that don't identify you.
• Site measurement (Section 5): the visit, download and installation records are kept for up to 400 days and then deleted. The record of which download and installation led to your account is kept with the account and deleted with it.
• Account data: kept while your account is active. On deletion, removed within 30 days, except a limited set of records we are legally required to keep, such as payment and tax records for up to 7 years under tax law, and minimal security records to enforce our Terms or defend legal claims.
• Consent records (when you turn data collection on or off, accept Terms or this policy, etc.): retained for the lifetime of your account, and for up to 7 years after account deletion, so we can demonstrate the lawfulness of past data collection if asked by a regulator.
• Feedback you submit: kept up to 3 years for product-improvement analysis, then deleted. Contact details attached to feedback are deleted sooner on request. Answers to the optional uninstall survey are kept up to 12 months, then deleted.
• Diagnostics from the desktop app (crash reports, update diagnostics, and uninstall notices): kept up to 90 days. After that we keep only aggregated counts that are not tied to any installation.
• Server logs: our application logs and our hosting provider's request logs are kept on a short rolling schedule (typically two weeks or less) and then automatically deleted. Where feasible, we remove the last segment of any IP address that our application records in logs so it cannot be traced to a specific household; some logs at the infrastructure layer (for example, DDoS-mitigation logs at our edge provider) may retain the full IP for short periods because that is operationally necessary.
• Backups: encrypted database backups are kept on a rolling 30-day schedule. When you delete data from your account, the data is also removed from production immediately, but a copy may still exist in older backups for up to 30 days while those backups age out of the window. During that period the data in backups is locked: it is not accessible to anyone except for disaster-recovery operations (e.g., restoring service after a database failure), is not used for any other purpose, and is not restored to production unless a full system recovery is needed. Once a backup expires, the deleted data is permanently gone with it.

8. Your rights

Depending on where you live, you have most or all of the following rights regarding your personal data. You can use the controls in Settings › Privacy for the most common ones, or email info@unitra.ai for anything else.

• Access: get a copy of the personal data we hold about you.
• Correct it: have inaccurate personal data fixed.
• Delete it: ask us to delete your account and associated personal data, subject to the limited legal-retention exceptions described in Section 7.
• Take it with you: receive your data in a structured, machine-readable format (data portability).
• Restrict processing: ask us to limit how we use your data while a question is being resolved.
• Object: object to processing we do under "legitimate interest." For the desktop app's crash reports, update diagnostics, session statistics, and uninstall notice, the switches in Settings › Privacy do this directly.
• Withdraw consent: for the optional Help Improve Unitra collection (which covers text, audio and screen samples), and separately for news and offers email, at any time. Withdrawing one does not affect the other. New collection stops as soon as our systems propagate the change, typically within minutes. A request already being processed when you withdraw may complete and be stored, but no further samples will be taken. Screen samples still waiting on your device are deleted when you withdraw. You can separately ask us to delete what we already collected.

We respond within one month of receiving a verifiable request. For complex requests we may extend this by up to two more months and will tell you within the first month if that applies.

If you believe your rights have been violated, you can complain to your local data protection authority, for example your national Data Protection Authority in the EEA, the Information Commissioner's Office (ICO) in the UK, or the Federal Data Protection and Information Commissioner (FDPIC) in Switzerland.

Your US state privacy rights

If you are a resident of California, Virginia, Colorado, Connecticut, or another US state with a comprehensive privacy law, you have additional rights including the right to know, delete, correct, port, and opt out of data collection for model training (use the Help Improve Unitra toggle). You also have the right to non-discrimination for exercising your privacy rights.

We do not "sell" your personal information as that term is defined under California law. Reporting the outcome of an ad click back to the platform that served the ad (Section 5) may count as "sharing" under that law; we treat the Global Privacy Control (GPC) browser signal as an opt-out from it, and you can also opt out as described in Section 5. We do not use your data for cross-context behavioural advertising beyond that report.

To exercise these rights, use Settings › Privacy or email info@unitra.ai. We respond within 45 days as required by applicable state law, with one possible 45-day extension where permitted.

Virginia residents have additional rights to opt out of targeted advertising, sale, and certain forms of profiling. We do not sell data or profile you; the only advertising-related processing is the conversion report described in Section 5, which the GPC signal or the Section 5 opt-out switches off. If we decline a request, you may appeal by replying to our decision; we will respond to the appeal within 60 days.

Voice samples and biometric law (Illinois BIPA and similar)

If you opt in to audio quality improvement, voice samples are collected solely for evaluating and improving speech-recognition accuracy. We commit that:

• Voice samples are not used to identify individuals.
• We do not create or store voiceprints.
• We do not generate persistent voice embeddings (mathematical fingerprints of someone's voice) that could be used to recognise a speaker across recordings. Where intermediate numerical representations exist transiently inside the speech-recognition model, whether on your device or, under cloud assist, on our private GPU server, they are computed in memory only and are not persisted, exported, or used for identification.
• We do not share voice samples or any derived data with third parties for biometric-identification purposes.
• Audio samples are linked to your account for at most 90 days, after which we permanently sever that link. In de-identified form they are kept solely to improve speech-recognition accuracy and are permanently deleted no later than 2 years after collection, well within the destruction timeline required by biometric-privacy laws such as Illinois BIPA (which permit retention until the collection purpose is satisfied, and in any event no more than three years after your last interaction). You can delete your own samples at any time during the first 90 days.

Illinois residents: if you would like to allow translation-text contributions but opt out of audio collection specifically, email info@unitra.ai and we will configure that for your account. (You can also turn off the unified Help Improve Unitra switch to stop both at once.)

9. Automated decisions

Unitra does not make decisions about you based solely on automated processing that produce legal effects or similarly significant effects on you. Subscription limits and rate limits are applied uniformly based on your tier and are not based on profiling.

10. Security

We protect your data with industry-standard technical and organisational measures, including:

• All data sent between your device and Unitra is encrypted in transit (the same "https://" lock you see in browsers).
• Translation text you opt in to share is encrypted before it is written to our database. The keys used to encrypt it are stored separately, and only a small set of authorised production systems can use them. The text recognised in screen samples is protected the same way, and screen sample images are encrypted on your device until upload and stored encrypted.
• Your password is stored using a one-way encryption method that is intentionally slow on purpose. Even in the worst case where our database were compromised, your original password could not be recovered.
• Login tokens expire after a short time and we can revoke them at any moment if needed.
• Database connections are encrypted; only internal services on private networks can connect to the database.
• Cloudflare protects our infrastructure from denial-of-service attacks and malicious traffic.
• We rate-limit suspicious activity, log administrative access, and follow the principle of giving each system only the access it strictly needs.

No security control is perfect. We continuously improve our program and welcome responsible disclosures at info@unitra.ai.

If something goes wrong

If we discover a personal data breach, we will notify the competent supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of it. If the breach is likely to result in a high risk to your rights, we will also notify affected users without undue delay, within the timelines required by applicable law.

11. Children's privacy

Unitra is a general-audience service. It is not designed for, marketed to, or directed at children, and we do not knowingly collect personal information from anyone below the applicable minimum age: generally 13, or the higher minimum required where you live (for example, 16 in much of the European Economic Area and 14 in certain jurisdictions).

We rely on the assumption that adults are creating their own accounts and on parents and guardians to supervise their children's use of internet services. If we become aware that an account belongs to someone below the applicable minimum age, we promptly delete the account and any associated personal data; the data is also removed from our backups as those backups age out of their retention window (see Section 7). Parents, guardians, or anyone aware of such a case can contact us at info@unitra.ai and we will act on the report without delay.

12. International data transfers

Our services and sub-processors are primarily located in the United States. If you access Unitra from outside the United States (including the European Economic Area, the United Kingdom, or Switzerland), your data will be transferred to and processed in the United States. The conversion reports described in Section 5 go to the advertising platform that served the ad (Google or Reddit, United States) and are made only for visitors outside the EEA, the UK and Switzerland.

Where required by law, we rely on one or more of the following safeguards:

• The EU–US Data Privacy Framework (DPF), its UK Extension, and Swiss–US DPF, where the recipient is certified.
• Standard Contractual Clauses approved by the European Commission, together with the UK International Data Transfer Addendum, where applicable.
• Additional technical and organisational protections (encryption in transit and at rest, access controls, and data minimisation) where the assessment of a particular transfer indicates more is needed.

You may request a copy of the relevant transfer mechanism for a specific sub-processor by emailing info@unitra.ai.

13. Changes to this policy

We may update this Privacy Policy from time to time. When we make a material change we will make the updated policy available in the app and on this page, and update the version dates at the top. Where a change expands the scope of processing that relies on your consent (for example, keeping voice samples longer), that new processing does not begin for you until you have reviewed the change and given a fresh, affirmative opt-in in the app. Until you do, we continue to operate under the choices you last confirmed, and any optional collection that depends on the new consent stays paused for your account. You can review, change, or withdraw your choices at any time in Settings › Privacy.

Screen samples (October 5, 2026). Adding screen samples to Help Improve Unitra (Section 3) is a change of this kind. If you turned on Help Improve Unitra before this date, a version of the app that takes screen samples will ask you to confirm it again. Until you do, no screen samples are taken from your account, and your text and audio contributions continue under the consent you already gave. If you turn Help Improve Unitra on in such a version, its description already includes screen samples.

14. Contact

If you have questions about this Privacy Policy or our data practices, or want to exercise any of the rights described above, contact us at:

Unitra, Inc.
Email: info@unitra.ai

We have appointed an internal Privacy Contact who oversees compliance with this policy and handles data-subject requests; the contact above reaches them.

EEA / UK / Swiss residents: if we are required by law to designate a representative in the European Economic Area, the United Kingdom, or Switzerland, we will appoint one and publish their contact details here. In the meantime, please use the privacy email above; we will respond within the same one-month timeframe required of EU-established controllers, and you retain the right to complain to your local data protection authority. We will also formally appoint a Data Protection Officer and publish their contact details here if and when that obligation applies to us.

Legal references

For data protection professionals, regulators, and others who want the underlying statutory references, the table below maps the sections above to the laws they implement.