The words vendors use for this

Half of these mean different things to different vendors, which is exactly the problem when you are comparing them. Where a term is slippery, this says so.

Localisation

Making content work in another language and context, not just translating the words. For video it includes the audio, the subtitles and anything written on screen.

Voice preservation

A translated voice intended to resemble the original speaker. Vendors mean very different things by this, from a generic voice with a similar tone to a model built from that person’s recordings. Ask which one, and ask what consent it requires.

Voice cloning

Building a synthetic voice from recordings of a real person. Whatever a vendor calls it, the consent question is the same: did that person agree to this specific use.

Terminology lock

Forcing a set of terms to be translated the same way every time. Useful when product names, safety terms or regulated wording must not drift between modules.

Sign-off

A named person approving a language version before it is published. The word is used loosely. Ask whether it is enforced by the tool or just a habit your team has.

Audit trail

A record of who changed what and when. Ask what is actually recorded and for how long, and whether you can export it.

Data residency

Where your content is stored and processed. Ask for the specific region, and ask where the AI providers in the chain process it too, because that is where the answer usually gets complicated.

Deployer

Under the EU AI Act, the organisation that uses an AI system and publishes the result. If you publish translated video, that is you, and some transparency duties attach to you rather than to your vendor.

Provider

Under the EU AI Act, the organisation that develops an AI system and puts it on the market. Provider duties do not discharge deployer duties.

Deepfake

AI-generated or manipulated content resembling real people, places or events that would falsely appear authentic. It is a legal category with disclosure consequences, not an insult.

Machine-readable marking

Embedding a signal in a file that identifies it as AI-generated, so that software can detect it. Distinct from a label a human viewer can see.

Burnt-in subtitles

Subtitles rendered permanently into the video image. They cannot be switched off or translated later, which is sometimes the point and sometimes a trap.

SRT and VTT

Standard subtitle file formats kept alongside the video, so subtitles stay switchable and editable.

Multi-speaker

Handling video with more than one voice, keeping each translated line with the right speaker. Interviews and panels break tools that assume one narrator.

Word emphasis

Marking words the translated voice should stress. It matters most for warnings, names and numbers, where a flat reading loses the meaning.

Speaking rate normalisation

Adjusting pacing because languages do not run to the same length. Without it, the translated audio drifts out of step with what is on screen.