Detection · audio deepfakes

Deepfake audio detection, and the distinction most tools skip

In short

A deepfake audio detector examines a recording for evidence that the speech in it was produced by a model rather than by a person. It cannot tell you whether a clip is dishonest. A recording can be entirely genuine and still deceive, and that failure is one no detector is built to catch.

The word deepfake gets applied to two problems that have almost nothing in common technically, share a name only because they share a purpose, and need completely different responses. Conflating them is how people end up trusting a green result they should not have trusted.

Two problems wearing the same name

The first is synthetic speech. Nobody spoke. A model produced the waveform from text, or from a target voice and a script, and the audio never existed as sound in a room until someone played it. There is a machine in the chain, and machines leave marks.

The second is manipulated real audio. Someone did speak, into a real microphone, and every sample in the file is authentic. The deception was manufactured afterwards, by cutting a qualifying clause out of a sentence, by joining two answers to different questions, by dropping a recording made in one week beside a claim about another, or simply by starting the clip a few seconds late. Nothing in that file is fake. The impression it produces is.

A detector answers the first question well and the second question not at all. Truthring is explicit about that boundary rather than blurring it, because the blurring is what makes results dangerous.

	Synthetic speech	Manipulated real audio
What happened	A model generated the waveform	A person spoke; the file was cut or arranged
Is the audio genuine?	No	Yes, every sample of it
What a detector reads	Traces of generation and of a missing recording chain	Traces of editing, which honest audio has too
Can Truthring answer it?	Yes, with a stated error rate	No
What resolves it instead	Analysis of the file	Provenance, the unedited source, the surrounding context

A worked example. An executive is recorded saying “if the numbers hold, and I doubt they will, we can promise a bonus.” Cut eleven words and you have a promise of a bonus, in their real voice, on genuine audio, from a real meeting. Every detector on the market would return human, correctly, and every one of them would be useless to the person trying to work out what actually happened.

How the detectable half is detected

Synthesis has to solve a perceptual problem: produce something a listener accepts. It does not have to solve a physical one. Nothing obliges a generator to model the acoustics of the room it was never in, the resonance of a chest that does not exist, or the specific hiss of the microphone that was never switched on. Those absences, and the statistical regularities a generator introduces in their place, are the material.

Truthring reads that material in layers rather than as one score. One layer asks whether the file behaves like something a physical sensor produced. Another compares fine-grained regularities against signatures collected from known systems, which is what allows a result to name a generator. A third watches how the voice handles the parts of speech that are hardest to fake convincingly: interruptions, restarts, laughter, a sentence abandoned halfway. Generators are trained on clean read speech and are least persuasive where real conversation is at its messiest.

The layers are described properly, with what each contributes, on the methodology page.

Where this breaks down

Four conditions push a result towards worthless, and the first two account for most real submissions.

The file has been through a network. Voice codecs are built to discard anything the ear will not miss, which overlaps almost exactly with what the analysis reads. A clip that has crossed a phone call and two messaging apps has been stripped three times.
The clip was captured from a speaker. Generated audio played into a room and recorded on a handset acquired a real recording chain honestly. The first layer now reports on a physical capture that genuinely occurred.
The generator postdates our most recent training run. The verdict often survives this; attribution usually does not, and drops to unknown generator rather than to a guess.
The deception was editorial. Covered above, and worth repeating because it is the failure people are least prepared for. There is no verdict here to get right or wrong.

Every known limitation, including the ones we have not yet quantified, is listed on limitations.

How to read what comes back

Three outcomes exist and they do not carry equal weight. Treating them as a symmetrical scale is the most common way a correct result gets misused.

Likely synthetic

The strong result. Something positive was found — a signature match, a recording chain that does not hold together. Read the confidence figure with it, and check the false positive rate for the condition your audio was in before acting.

Likely human

The weak result, and the one most often over-read. It means no evidence of generation was found, which on a compressed or short clip may simply mean the evidence did not survive. Absence of a finding is not a clearance.

Unclear

A real answer, not a failure. The audio did not support a judgement. Collapsing this into “human” would manufacture reassurance out of nothing, so it is returned as a third outcome in its own right, and never billed.

None of the three is a finding about a person. A verdict describes a file. Whoever sent you that file may have made it, may have been sent it themselves, or may have recorded it from a television. If a decision about someone’s money, job, liberty or safety is riding on the answer, the result belongs in the evidence pile alongside everything else, not on top of it.

Questions

What counts as deepfake audio?

In ordinary use, any recording engineered to make a listener believe something false about who spoke or what was said. That covers wholly generated speech, a cloned voice reading a script, and a genuine recording cut so that the meaning inverts. Only the first two are things a detector is built to find.

Can a deepfake audio detector spot edited recordings?

Not reliably, and Truthring does not claim to. Cuts, splices and reordering leave traces in the waveform, but so does every legitimate edit ever made to a podcast or a news package. Editing is evidence of editing, not of dishonesty, and separating the two is a job for context rather than for a model.

Is a synthetic verdict enough to call something a deepfake?

No. It says the speech in the file was probably manufactured, which is also true of an audiobook, a voiceover and an accessibility tool reading a page aloud. Deception is a claim about intent and circumstance. A detector reports on the audio and stops there.

Why do detectors do so badly on viral clips?

Because by the time a clip is viral it has been through several apps, each of which re-encoded it, and often it has been screen-recorded from a video. What arrives is a copy of a copy. Most of the fine structure a detector reads has already been thrown away in the interest of file size.

What should I do with an audio clip I think is fake?

Keep the earliest copy you can obtain and stop forwarding it, because every hop degrades it. Establish where it came from before you analyse it: a clip with no traceable origin is a weak exhibit whatever a detector says about it. Then treat the verdict as one line of evidence among several.