Detecting LOVO
LOVO is a commercial speech company whose output sits mainly in marketing and e-learning — the thirty-second radio spot, the product demo, the onboarding module a company sends every new starter. When someone checks a LOVO clip they are usually not a victim. They are a buyer, a broadcaster, a regulator or a performer, and each of them wants a different thing from the same verdict.
Measured on [VERIFY: n] clips spanning [VERIFY: n] voices and [VERIFY: n] read styles, generated on [VERIFY: date] and held out of training. Method: accuracy.
Four people, one clip, four questions
Detection on a fraud page has one customer. Here it has several, and they are not asking the same thing.
The brand
Bought a campaign from an agency, was invoiced for a performer, and wants to know what was actually delivered. The useful output is evidence for a conversation with a supplier, not a public claim.
The broadcaster or platform
Has a disclosure policy and needs to apply it consistently across thousands of spots. Wants a repeatable procedure with a known error rate far more than it wants a single dramatic result.
The performer
Recognises something close to their own delivery in an advert they never recorded. Needs to establish that the audio was generated before any question about likeness or licensing can even begin.
The compliance reviewer
Is checking whether a required disclosure was made. Their question is about a label somewhere in the creative, and the audio only supplies half of it.
Why commercial voiceover is a favourable case, right up until it is not
Advertising audio starts out easy for us. A voiceover is recorded or rendered to a clean master, is normally long enough to give the analysis real material, and is delivered as a high-quality file. Read speech at an even level is exactly the condition under which a generator has the least reason to fake the physical evidence of a recording session.
Then post-production happens. A spot is compressed for loudness, ducked against music, layered with effects, matched to a broadcast target and finally encoded for whatever medium carries it. Each of those steps is designed to make the voice sit better in a mix, and each incidentally removes the fine spectral detail the analysis reads. By the time you capture a radio ad off air, a substantial part of the evidence has been processed away — which is why the figures above fall so steeply across the rows.
The practical rule follows directly: get as close to the delivered master as you can. A brand chasing a supplier usually can, because the master is a deliverable they paid for. A member of the public complaining about a spot usually cannot, and their result should be read with that in mind.
Disclosure, carefully
Rules requiring advertisers to disclose synthetic media are expanding, and they differ by market, by medium and by what is being synthesised. We are not going to summarise them here, because a confident summary that is wrong in your jurisdiction is worse than none. [VERIFY: verify the rules that apply in your market and link them.]
What is stable enough to say is the shape of the concern. Regulators have been markedly more interested in synthetic speech that imitates a specific identifiable person — a celebrity endorsement nobody gave, a public figure appearing to recommend something — than in a generic narrator that never claimed to be anyone. A stock synthetic voice reading ad copy is, in most places, a labelling question at worst. A synthetic voice that sounds like a named individual is a different order of problem, and one where the audio evidence is only the opening move.
There is also a contractual layer that has nothing to do with regulators. Performer agreements increasingly say something about whether a recorded session may be used to train or generate. A detection result can establish that delivered audio was generated. It cannot tell you what the contract said, and the contract is where the dispute will actually be decided.
What each verdict is worth here. Likely synthetic is the stronger result, because it is only reached on positive evidence in the file. Likely human is weaker, and on heavily processed advertising audio it is weaker still — it can mean a person read the script, or it can mean the mix and the encoder removed what we would have read. Do not treat a clean bill of health on an off-air capture as a finding.
If you are checking a campaign
- Ask the supplier for the delivered master and the voiceover stem, not the finished mix. The stem is the single biggest improvement available to you.
- Check several spots from the same campaign. A consistent reading across a set is much harder to argue with than one result on one file.
- Keep the reference code from each result. It lets a supplier, or their lawyer, reproduce and challenge the analysis, which is the point.
- Put the result to the supplier as a question before you treat it as a finding. Agencies sometimes do not know what their own subcontractors used.
- If a specific performer’s likeness is involved, stop treating this as a detection problem. It becomes a rights problem, and the audio is one exhibit in it.
Questions
Can LOVO voiceover be detected?
On a delivered master, Truthring reads it as synthetic in a share of held-out clips that has not been measured yet and names LOVO specifically in a rate not yet measured of those. Both figures fall through the post-production chain, and attribution falls first.
Does an advert with synthetic voiceover have to be labelled?
It depends on where it runs and what the voice is doing. [VERIFY: verify the applicable rules.] The recurring theme across markets is that imitating an identifiable person is treated far more seriously than using a generic synthetic narrator.
An agency invoiced us for a voice session. Can you prove they did not run one?
No, and be careful with that framing. We can say whether the delivered audio reads as generated. A session may have happened and been discarded, or a human read may have been processed until it reads oddly. The verdict opens the conversation; the paperwork closes it.
I think an advert copies my voice. What do I do first?
Preserve the highest-quality copy of the spot you can obtain and note where and when it ran, before you contact anyone. Then get a reading on it. Establishing that the audio was generated is the step that makes every later question — licensing, likeness, contract — worth asking.
Why do two spots from one campaign score differently?
Different mixes, different music beds, different encodes. The underlying narration can be identical and still produce different confidence figures once it has been through different post-production. Compare readings from files at the same stage of the chain.
Signature last retested [VERIFY: date] against LOVO model version [VERIFY: verify]. Rates on this page are re-measured monthly and move when the vendor ships.
Reviewed