Audio verification, not audio classification
Audio verification establishes what can defensibly be said about a recording and puts it in a form a third party can check. A classifier hands back a label. A verification hands back a stated method, a measured error rate, a model version and a reference code that lets the whole thing be examined again.
The difference matters exactly once, and it matters completely: at the point where somebody refuses to accept your answer. Until then a percentage is fine. After that, a percentage on its own is worth nothing at all.
A label is not a finding
Suppose a tool tells you a recording is 94% likely to be synthetic. Now suppose the person it concerns says it is not. What do you have?
You have a number with no provenance. You cannot say which model produced it, what that model’s error rate is on audio that has been through a phone network, what would have to be true for the number to be wrong, or how anyone would obtain it again in six weeks when the tool has been retrained twice. The number was never the finding. It was one input to a finding that nobody wrote down.
That gap is why Truthring is built around a report rather than around a score. The analysis is the same either way. What changes is whether the output can be handed to a sceptic and still hold together.
The five stages, and what goes wrong at each
Most of the value sits in the two stages people skip, which are the third and the fourth.
| Stage | What happens | What you get | Where it goes wrong |
|---|---|---|---|
| 1. Upload | The earliest available copy of the file is submitted, with the channel it came from declared. | A clip in a known condition, and a record of that condition. | Submitting a version that has been forwarded through three apps. Each hop removes evidence permanently. |
| 2. Analyse | Layered examination of the recording chain, the generator signatures and the behaviour of the voice. | A verdict, a strength figure, and the name of a generating system where one can be given. | Nothing, usually. This is the part everybody already does, and the part that is least often the weak link. |
| 3. Interpret | The result is read against the false positive rate measured for audio in that condition. | An honest sense of what the verdict is worth on this specific file. | Reading a confident number without the error rate beside it, or treating an absence of evidence as a clearance. |
| 4. Verify | The conclusion is checked against everything outside the audio: where the file came from, who held it, whether the account it is attached to holds up. | A finding rather than a measurement. | Skipping it. Origin is what an opponent attacks first, and a file that cannot be traced anywhere stays thin whatever the analysis returned. |
| 5. Share | The whole chain is written into a report with a reference code and issued. | Something a third party can read, reproduce and argue with. | Circulating a screenshot of a percentage, which strips out every part that made it defensible. |
What makes a result survive disagreement
Four fields carry almost all of the weight. They are unglamorous, none of them is the verdict, and a result missing any of them collapses under the first serious question put to it.
- Stated method
- What was measured and in what order, written so that a reader who did not run the analysis can follow it. Published in full at methodology rather than summarised as “proprietary AI”.
- Stated error rate
- How often authentic speech gets mistakenly marked as generated, measured on audio in the condition this clip arrived in — not one headline figure taken from clean studio files. [VERIFY: measured false positive rates pending first benchmark]
- Model version
- The exact version that produced the verdict, stored beside it. Detection models are retrained; a conclusion that cannot name its own version cannot be reproduced after the next release.
- Reference code
- A short identifier that resolves to the analysis, so a person who was not in the room can pull it up and disagree with it in specific terms.
A worked example of the whole document: specimen report TR-8F29A1.
Who actually needs this
Not everyone, and saying otherwise would be selling. If you have received an odd voice note and want to know whether to worry, run it through the free check, ring the person back on a number you already had, and get on with your day. No report is required for that and none should be sold to you.
Verification starts to matter when the answer will be contested by somebody with an interest in contesting it. A grievance where a recording is the central exhibit. An insurance claim resting on a recorded instruction. A newsroom deciding whether to publish, knowing the subject will respond. A payment dispute where a bank wants to see how the conclusion was reached, not merely what it was. In each case the question is not what does the tool say but will this hold when someone attacks it.
The same discipline is also what keeps us honest. Publishing the method, the measured error and the version is uncomfortable in a market where competitors publish one flattering figure. It is also the only arrangement under which a customer can tell whether we are any good.
Verification is being built now
Truthring has not launched. Reports, references and the API are design commitments rather than shipping features, and we would rather describe them accurately than imply they are ready.
Questions
What is audio verification?
The process of establishing what can responsibly be said about a recording and recording it in a form somebody else can examine. Classification is one step inside it. Verification adds the method, the measured error rate, the model version and a reference, so the conclusion can be tested rather than simply believed.
How is that different from an AI voice detector?
A detector produces a label. Verification produces a document. The label is the same in both cases; what changes is whether a third party who was not present can reconstruct how it was reached, see the conditions it was reached under, and challenge it on the record.
Do I need verification, or is a classification enough?
If nobody is going to question the answer, a classification is enough and you should not pay for anything more. Verification earns its cost at the moment someone disputes the result: a disciplinary hearing, an insurance claim, a court bundle, a correction request to a newsroom.
Can a verification report be used as evidence?
It can be submitted as one exhibit among others, and it is built to survive being read by someone hostile to it. It is not proof and nothing on this site will describe it as proof. A probability with an honest error rate is evidence; a probability presented as a fact is a liability.
What happens to a report after the model is retrained?
It stays valid as a record of what was concluded, when, and by which version. The reference code resolves to the original analysis with its original model version attached, and the method changelog records what changed afterwards. A result you relied on in one month remains explicable in the next.
The evidence chain
Detection pages
Reviewed