Voice is the modality that arrived first and got least attention, probably because nobody makes a showreel out of a voiceover. It is also the one where the quality argument is effectively over for a large category of work, which moves the interesting questions elsewhere.
Where it passes and where it does not
| COPY TYPE | PASSES? | WHY |
|---|---|---|
| Neutral informational VO | Yes | Even pace, no stress decisions, no subtext to carry |
| Corporate narration | Yes | The register is already flat, which is a low bar to clear |
| E-learning and instruction | Yes | Clarity is the whole requirement |
| Emotional read | Rarely | Emotion is produced by decisions mid-sentence, which is exactly the gap |
| Comedy | No | Timing is the mechanism, and it does not survive synthesis |
| Character performance | No | Listening, hesitation and self-interruption are the craft |
| Anything with a name on it | Depends entirely on consent | A rights question before a quality one |
The six controls that make it sound human
Most bad synthetic reads are undirected. The model was capable. The controls that matter, in order of effect:
- Breath placement. Insert breaths where a person would take them, which is before a clause they are about to emphasise. This single change carries the read.
- Pace variation. A human read speeds up through the familiar and slows through the important. A uniform pace is the loudest tell there is.
- Emphasis by rewriting. Move the word you want stressed to the end of the clause. Prosody follows structure. Tags are a suggestion.
- Sentence length variation in the script. A synthetic read puts monotonous rhythm on full display, because a human voice varies against it without being asked to.
- One imperfection. A slightly early breath, a very small stumble, one word taken at a different pace. One only.
- Room. A clean voice in a silent mix is not a recording of anything. Put it in a space.
The consent structure
A cloned voice needs a release that a standard performance contract does not contain, because the standard contract licenses a recording and this licenses a capability.
- Explicit grant to create a model from the supplied recordings, named as such.
- Scope: which brands, which product categories, which media. A voice licensed for internal training that appears in an advert is a breach nobody documented.
- Term, with a stated end date and a defined disposal obligation for the model when it arrives.
- Territory, which matters because personality and likeness rights differ sharply between jurisdictions.
- Exclusions: categories the performer will not be used for. Political, gambling, alcohol, health claims. Whatever they choose.
- Withdrawal: a mechanism, a notice period, and what happens to assets already in market.
- Rate structure for use, as well as for the session. A session fee for something that runs forever is the arrangement performers are right to resist.
Disclosure
A wholly synthetic voice presented as nobody in particular is, in most markets, outside the disclosure duty, though platform policies vary and frequently reach past the law.
A cloned voice of an identifiable person is a different matter. Since August 2026, EU transparency obligations under Article 50 apply to synthetic audio that qualifies as a deepfake, and an identifiable person’s cloned voice sits squarely in that. The working position is to disclose, to hold the consent in writing, and to decide both at brief stage.
What we will not do
Clone a voice without a signed release from the person, including for an internal test, and including when the recordings are sitting in public. The availability of material has never been the same thing as permission, and the fact that it is now technically trivial is an argument for the rule.
The definition, the consent structure it requires, and the related terms.
VOICE CLONING, DEFINED →