Audio provenance

The idea that files should carry their own history, so nobody has to reconstruct it later. It is the right direction of travel and it has one structural flaw that will not go away.

Three mechanisms, three different weak points

Signing. A capture device or generation tool attaches a cryptographic assertion — this file was produced here, at this time, in this way — which can be checked later and which breaks if the file is altered. Strong where it exists. Its weakness is that it is a claim by whoever held the signing key, and it says nothing about what was in front of the microphone.

Watermarking. A signal deliberately embedded in the audio itself, ideally inaudible, ideally surviving compression and re-encoding. Its appeal is that it travels with the sound rather than with the container. Its weakness is that survival is a design trade-off: a watermark robust enough to survive a phone codec has to be strong enough to risk being audible, or to be removed by someone who knows it is there.

Container metadata. The device, app, encoder and timestamps recorded alongside the audio. Enormously useful in practice and trivially lost: exporting a voice note from a chat thread frequently discards all of it, and metadata is not authenticated in the first place.

The flaw that does not go away

Provenance is a positive signal. Its presence tells you something; its absence tells you almost nothing.

A file without credentials might be a fabrication. It might equally be a genuine voicemail from a phone that has no signing capability, a recording made a decade ago, a clip legitimately re-encoded on its way through three apps, or audio from anyone who never opted in. Since unsigned material will remain the overwhelming majority for the foreseeable future, a system that treats missing provenance as suspicion will spend its time flagging ordinary people.

The adversarial version is simpler still. Signing and watermarking are voluntary at the point of generation. Someone building a clone for fraud will choose a tool that does not mark its output, or run a model locally where nothing is enforced. Provenance raises the floor for honest publishers and does not raise the ceiling against a motivated attacker — which is exactly why detection remains necessary, and why the two are complements rather than alternatives.

A concrete case

A woman keeps a threatening voicemail. It carries no credentials of any kind, because voicemail systems do not issue them. Two years from now, in a world where signed media is common, the absence has acquired a meaning it does not deserve: the other side argues that genuine audio would have been signed.

Nothing about her recording changed. The expectation around it did. That is the mechanism by which a provenance ecosystem, deployed carelessly, hands a new argument to the person our page on the liar’s dividend describes.

Commonly confused with: audio forensics

They answer the same question from opposite ends of the timeline. Forensics reconstructs a file’s history after the fact, from whatever traces survive — encoder structure, edit boundaries, room acoustics, chain of custody. Provenance tries to record that history at the outset so the reconstruction is unnecessary.

Provenance is cheaper and more reliable when it is there. Forensics is what you have when it is not, which today is nearly always. Neither asks whether the speech was generated: a signed file can contain a clone, correctly signed by the tool that produced it.

Where Truthring sits relative to it

We analyse the signal rather than any accompanying claim, because the files people bring us — forwarded voice notes, voicemail exports, screen-recorded clips — have usually lost every credential they might once have had. What we do add is a small piece of provenance at our own end: a one-way hash of the submitted file is retained after the audio is deleted, so a report can be tied back to the exact bytes it described, and the engine version is stamped on it so it can be read against the method changelog.

[VERIFY: which provenance and content-credential standards Truthring will read, verify or emit, if any, and from what release]

If provenance becomes widespread, the honest position will be the same one we take about our own output: a credential is evidence, not proof, and its absence is not evidence of anything at all. Our transparency notes hold that line.

Questions this term raises