Reading your result
A result card carries five pieces of information and one large caveat. Most misreadings come from taking one of the five in isolation.
Likely fabricated
Below the headline sits a plain-language reason, then the four rows. This card is an illustration of the layout, not a real analysis.
- Attributed model
- [VERIFY: named system, or Unknown generator]
- Source quality
- [VERIFY: e.g. Phone-compressed · 0:14]
- Audio
- Deleted · hash kept
- Ringmark
- [VERIFY: e.g. TR-XXXXXX]
The verdict
Three states are possible, and they are not symmetrical. Read the chip and the headline together — they are two phrasings of the same finding.
Likely fabricated. Generation signatures were found and the recording carried enough usable signal to trust them. This is the stronger verdict, because it required positive evidence: something in the file matched something we hold. It does not establish who made the clip, when, or why.
Likely human. Nothing matched, and the audio behaves like captured speech. This is the weaker verdict, and the wording is deliberate. The same result appears when a generator we have never seen was used, and when the evidence of synthesis was destroyed in transit by forwarding, re-compression or noise reduction. Absence of a signature is not authenticity.
Not enough signal to answer. Too short, too compressed, too noisy, several overlapping speakers, or edited after the fact. This is a real answer rather than a failure, and it is the answer most detectors will not give you. The right response is usually a better recording, not a second opinion. How to get one.
Confidence
The large figure is a percentage attached to the verdict shown beside it. It is not a standalone likelihood that the audio is synthetic. When the verdict is unclear, no figure is shown at all — the card displays a dash, because a confidence number for a non-answer would be meaningless.
Confidence is capped by the quality of the recording before it is reported. A strong signature in a badly degraded file is deliberately reported as a weaker result, because that is what it is. This is why the same underlying clip can score differently depending on how it reached us.
Attributed model, source quality, ringmark
Attributed model names the generation system when the clip matches a signature we hold. When it does not, the row reads unknown generator. That phrase means we cannot name it — not that nothing was used. A named model is a useful corroborating detail; it is not a stronger form of the verdict.
Source quality describes what the file has been through and how much usable speech it contained. It is the single best predictor of whether a repeat analysis on a better copy would tell you more.
Ringmark is the reference code for that specific analysis. Quote it in any correspondence about the result. It ties the answer to the exact file and the engine version that produced it, which matters because the model is retrained as new generators appear. The audio itself is deleted once analysis completes; a one-way hash is kept so the link survives without the file. Privacy.
Four mistakes people make
- Reading the confidence without the verdict. A high number beside likely human and a high number beside likely fabricated point in opposite directions. Quoted alone, the percentage says nothing at all.
- Treating likely human as clearance. It is the weaker verdict by construction. It is a reason to keep verifying through a channel you trust, not a reason to stop.
- Over-reading the attributed model. A named system tells you what the signature resembled. It does not tell you who used it, and unknown generator is not a vote for authenticity.
- Treating the result as proof. It is an instrument reading with published limitations. For anything involving money, employment, legal claims or safety, it is one data point beside a callback on a number you already have.
Questions people ask
Does likely human mean the recording is genuine?
No. It means no generation signature was found in the file as it reached us. A forwarded, re-compressed or noise-reduced copy can lose every trace of synthesis and still return this verdict.
Why is there no confidence figure on my result?
Because the verdict was unclear. There is no answer for the number to qualify, so the card shows a dash rather than an invented figure.
Two copies of the same clip gave different answers. Why?
Almost always because the copies are not the same file. Each forward, re-encode or export changes what survives. Analyse the original wherever you still have it.
Can I cite this result somewhere formal?
A verified report carries the file hash, a timestamp, the engine version and the limitations that applied, and can be checked at a permanent URL. It remains a probability, and whoever reads it should be told that. Pricing.
[VERIFY: written by — real name, real role]. Reviewed [VERIFY: date]. Describes engine version [VERIFY: version]. Corrections: method@aivoicedetctor.com.
Reviewed