AI Voice Cloning and Dubbing

Voice is the most personal thing a generative system can reproduce and the easiest to misuse. The technology has been usable for a while; the part that determines whether a project is defensible is entirely about consent, verification and control.

The consent framework, which is the engagement

Before any model is trained, this is established and documented:

• Informed, specific consent. The speaker knows a voice model will be created, what it may be used for, in which languages and territories, and for how long. Blanket permission is not consent.
• Verified identity. That the person consenting is the speaker in the recordings, checked rather than assumed.
• Named prohibited uses. What the voice will never be used for — political content, endorsement, adult content, or anything the speaker specifies.
• Revocation that works. A defined process to withdraw consent, with the operational consequences understood by both sides before it is needed.
• Compensation terms. Whether use is per-project, per-use or licensed, agreed rather than left to be argued later.
• Retention and deletion. How long the model and the source recordings are kept, and what deletion actually removes.

What we will not do

1. Clone a voice without the speaker's documented consent. No exceptions for public figures, deceased individuals, or recordings that are publicly available.
2. Build a voice to impersonate someone in order to deceive a listener about who is speaking.
3. Produce content designed to defeat voice authentication or to support social engineering.
4. Generate a voice saying things the speaker refused, or that falls inside their named prohibited uses.
5. Omit disclosure where a listener would reasonably assume they were hearing a real recording of that person.

These are refused whatever the stated purpose, including for testing, demonstration or internal use. A voice model that exists can be misused later, so the control has to be at creation.

Building it well

Record for the purpose where you can

Purpose-made recordings with clean audio and phonetic coverage produce noticeably better results than assembling material from existing content. Where only archive material exists, we say what quality is achievable from it before committing.

Direct the performance, not just the words

Pace, emphasis, warmth and pauses are what make synthetic speech sound intentional rather than merely correct. Script markup and per-line direction are part of the workflow, and they are the difference between usable and uncanny.

Handle the things that break it

Names, technical terms, numbers, acronyms and brand pronunciations need a pronunciation dictionary. This is unglamorous and it is what stops a polished production being undermined by a mispronounced product name.

Dub for meaning, not word count

Translation for dubbing is adaptation — timing, register and cultural fit — and it needs native review. Direct translation produces dubs that are technically accurate and obviously wrong to a native speaker. See AI translation and localisation.

Watermark and log every generation

Audio watermarking where the tooling supports it, plus a log of what was generated, from which model version, by whom and for what. This is how a misuse question is answered later, and it is also what makes revocation enforceable.

Keep a human approval gate

Nothing in a real person's voice publishes without a named human approving that specific output. Generation at volume without an approval step is how a voice model produces something nobody intended.