AI Transparency Notice

Everything you read here is machine-generated

SubLive has no human in the loop. Speech is turned into text by a speech-recognition model, and that text is translated by a language model. Nobody reviews the result before you see it, and there is no version of the product where somebody does.

Subtitles are marked as AI-generated in the interface itself, not only on this page, because that is the moment it matters (EU AI Act, Article 50).

How it fails

The failures worth knowing about are the ones that do not look like failures:

• Confident wrong words. Speech recognition does not return "I did not catch that" — it returns its best guess, in fluent, plausible language. Names, numbers and technical terms are where this happens most.
• Fluent wrong meaning. A translation can be grammatical and natural and still say something the speaker did not say, particularly with negation, irony, conditionals and figures of speech.
• Lost register. Politeness, hedging and formality often do not survive translation. Something said carefully can arrive blunt, and the reverse.
• Missing context. The model sees the last few utterances, not the whole conversation, and not what is on your screen.

What it is not for

This is a comprehension aid. It is not a certified, sworn or legally binding translation, and it is not a record of what was said.

Do not rely on it alone for anything you cannot undo: signing, agreeing to terms, medical or legal decisions, or judging a candidate. For those, use a qualified human interpreter or translator. If a subtitle matters, ask the person to repeat or confirm it — that costs a sentence and settles it.

The spoken voice is synthetic

If you switch dubbing on, the translation is also read aloud. That voice is not a person and is not the speaker: it is generated, either by your own browser or by a speech synthesis model, and it belongs to nobody. It is picked from a fixed set of neutral voices, and different speakers get different ones only so you can tell them apart.

We do not clone the voice of anyone in the conversation, and we do not build voiceprints. The speaker's own voice is never used to produce the dubbing — only the translated text is.

The dubbed audio plays to you and stops there. It is not routed back into the call, and nobody on the other side hears it. If you want the people you are talking to to know a machine is translating, that remains yours to tell them (EU AI Act, Article 50).

The dubbing carries every failure listed above, plus one of its own: it lags behind the speaker and, on continuous speech, it can skip ahead and leave a line unspoken. The subtitles are the complete record of what was translated; the voice is a convenience on top of them.

Summaries of saved conversations

If you saved a conversation, you can ask for a summary of it. The summary is written by a language model that reads the recognised speech — not the audio — and it inherits every failure above: a misheard number stays misheard, and a sentence the model did not understand can come back as a decision nobody made.

A summary is a reading aid, not minutes of a meeting and not a record of what was agreed. Check anything that matters against the transcript next to it, which is why both are shown on the same page.

Summaries are generated only when you ask for one, are stored with the conversation they belong to, and are deleted together with it — including when you delete the conversation yourself.

Asking for one can also be a standing request: if you switch on conversation recaps, a summary is written when a conversation ends and emailed to you. The email leaves our system for your mailbox, so deleting the conversation here does not delete the copy that already arrived. Recaps are off unless you turn them on.

When translation fails

If a translation does not arrive in time, we show you the original text and mark the line as degraded instead of showing nothing. Silence is the worse failure for a live conversation, but you should know that such a line was not translated at all.

Which models we use

The specific speech-recognition and translation providers are listed on the Subprocessors page and can change. We give notice before adding a provider, as described there.

Your speech is not used to train anyone’s models under our contracts with these providers. We do not build voiceprints and we do not identify speakers by voice — see the Biometric data section of the Privacy Policy.