Privacy & Security

1 · How data flows

Once you see how Oralion works, it's clear why your content is hard for us to reach. A broadcast involves four kinds of data, each on its own path:

2 · The speech provider's commitments

Recognition and translation are performed by a specialist real-time speech provider headquartered in the US, whose service and data handling are subject to applicable US laws and privacy standards. Privacy was central to our choice of partner. Per its privacy terms, the provider commits to:

• A US-based company: operated by a US-headquartered team, following US data-protection practices.
• Real-time processing, no long-term retention: audio and transcripts are used to return results in real time and not stored long-term.
• Not used to train AI: your audio and text are never used to train or improve any AI/ML model.
• No secondary use, no selling: your content is not used beyond delivering the service, nor sold or rented to third parties.
• Encrypted in transit: the connection to the provider is encrypted throughout.

These are the provider's commitments under its privacy policy. Oralion itself never handles this audio, so it cannot and does not retain the content.

3 · What we collect

We deliberately keep collection to a minimum. In practice it's only the following, none of which includes audio or caption content:

• Account data: your email, a hashed password (we can't recover your original), the stable identifier when you choose Google sign-in, and usage/credit stats. We do not store Google access tokens.
• Registration IP, rough city and browser identifier: each kept once at signup, solely to spot duplicate registrations and prevent abuse. From the third account created in one browser, the new account gets a smaller monthly free allowance; paid plans and purchased permanent Credits are unaffected. Visible to admins only, never made public or used for anything else.
• Room code and room password: held only in server memory to set up the live connection; they vanish when the broadcast ends, and are never written to the database or logs.
• Anonymous trial metering: logged-out trials are metered temporarily by source IP (in memory only) to grant the free allowance and prevent abuse.
• Broadcast session metadata: after a broadcast ends we keep a summary for operations and usage analysis — room code, language settings, consecutive/simultaneous mode, start/end times, and viewer count. Any geographic info is only a rough country/city (derived once from the IP, after which the raw IP is not kept), and it never includes audio or caption content.
• Cookies: three, all our own, none used for ad tracking. The session cookie that keeps you signed in; locale, which remembers the language you chose; and oralion_trial, a random identifier used to meter the signed-out free trial and to tell whether repeated accounts are being created in one browser (see section 3). It contains nothing personal and cannot be read by scripts on the page.

4 · What we never do

• We don't store your audio.
• We don't store your original or translated caption content — AI live correction relays it in real time and stores nothing either.
• We don't sell or rent your content or personal data to anyone.
• We don't use your data for advertising or cross-site tracking.
• We don't reuse your content in any form or use it for AI training.

5 · Security measures

• HTTPS/TLS everywhere: all pages and connections are encrypted in transit.
• End-to-end encryption: captions and translated voice are readable only at the speaker and viewer ends.
• Short-lived credentials: the credentials used to reach the speech service are short-lived and issued per session; our master key never leaves the server.
• Hashed passwords: account passwords are stored hashed with an industry-standard algorithm, never in plaintext.
• Server-enforced room passwords: for a password-protected room, an unauthenticated viewer can't obtain a connection at the server, so they never receive any captions.
• Proof links: a speaker can create a separate proof link for the proofreaders working with them. Anyone holding it can read and change that broadcast's captions, and changes appear to the audience at once; they are encrypted in the same way and our servers cannot read them. It uses a different key from the viewer link, and the speaker can invalidate it at any time.

No online service can guarantee absolute security, but we continually reduce risk through least-privilege and end-to-end-encryption principles.

6 · Retention & deletion

• Account data: kept until you delete your account or request deletion.
• Audio and captions: not retained by default. Raw microphone audio does not pass through us. Caption and translated-voice ciphertext is relayed in room memory only and never persisted. Full plaintext captions used for AI correction are not stored, but the terminology-rule audit records described above are retained. Optional transcript saving, and the automatic summaries produced from it (which send the transcript to a language model), are covered in section 8.
• Room code / password: cleared from memory when the broadcast ends.
• Session metadata and server logs: kept only for the limited period needed for operations analysis and abuse prevention.

7 · Third-party services

To provide the service we use a few third parties:

• Speech provider: real-time recognition and translation, plus any recognition context you choose to provide (see section 2 for commitments).
• Language-model provider: used to correct captions while AI live correction is on, or to organise optional recognition context when a signed-in speaker enters it. This is the only third party that sees caption text; turning live correction off stops captions from being sent. Neither path includes your account, email, or room code. When a speaker opens a question room, the same provider also translates audience questions and the asker's introduction (never the asker's name).
• Google: used only when you choose Google sign-in, to verify your identity and obtain your email and stable account identifier; access tokens are not retained.
• Cloud hosting: runs our website and signaling/relay servers.
• Email delivery: sends account confirmation and notification emails.
• Noise-reduction model provider: used only when a speaker turns on background noise reduction. The processing runs entirely inside the speaker's own browser and no audio is sent to this provider; the browser only downloads a model file from them, so they see the speaker's IP address.

8 · Transcript saving and summaries (opt-in)

Captions are not stored by default. Only after you explicitly enable transcript saving in account settings will a finished broadcast be saved to your own account for export or deletion. We do not use these transcripts to train anything, and we do not use them to evaluate our own quality. What is stored is used for one thing: producing the summaries described below, for you to look back at.

Automatic summaries

• So that you can find a particular session later, a transcript is sent to a language model to produce a title, the people who took part, and a summary of the main points. At most 20,000 characters per session; a longer one is sent as its beginning and its end, with the middle left out. Only one language of a transcript is sent. The text passes through that one request only; it is not retained by the model provider and not used for training.
• Transcripts held only on your device are summarised too, and the transcript itself is still never uploaded or stored. Your browser sends that same extract, we produce the summary, and only the summary is kept — the transcript stays in your browser. This is counter-intuitive, so plainly: the transcript is not kept, the summary is.
• Summaries are produced by a model and can be wrong or miss the point. They are an aid to remembering, not minutes of record.

Scope and limits

• Transcripts belong to your account and are kept until you delete them; closing your account deletes them too.
• You can turn it off at any time in account settings, or delete individual transcripts or all of them. Turning it off stops new broadcasts being saved; what is already stored stays until you delete it.
• A summary is deleted together with the session it belongs to, and everything is deleted when you close your account. For the ones held only on your device, deleting the summary leaves us with nothing at all.
• Categories you create on the transcripts page store their names, and which session is in which, in your account so every device shows the same ones; they hold no transcript text. A session's category goes when the session does, and all of them when you close your account.
• When you press “Auto-file”, the titles and main points of your unfiled sessions' summaries, together with your existing category names, are sent to a language model to choose categories. The transcripts themselves are not sent, and the model provider does not keep them or train on them.

Enable or disable transcript saving in account settings.

9 · Your rights

You can view, correct, or delete your account data at any time. Once logged in, update your email and password in account settings; to delete your account (including the Google sign-in identifier) or exercise other data rights, email [email protected] and we'll handle it within a reasonable time.

10 · Children's privacy

This service is not designed for children. We don't knowingly collect children's personal data; if you believe a child provided data without consent, contact us so we can delete it.

11 · Changes & contact

We may update this page from time to time; material changes will be marked with a new "Last updated" date here. Continued use of the service means you accept the updated terms.

Questions about privacy or security? Email [email protected].