SoundmarqueSoundmarquePartners
Provenance

The industry can say a recording used AI. It cannot yet prove who said it.

Every delivery format now carries a field for it. Charts have rules that turn on it. None of it is bound to the audio. The flag travels as metadata that anyone downstream can set, clear, or lose in a transcode, and nothing in the file says who put it there. It works exactly as long as everyone in the chain is honest and careful.

A declaration and a signed record are not the same object.

One is a field somebody filled in. The other is a statement with a name on it, bound to the audio it describes, timestamped by an authority that is not us, and checkable by anyone holding the file. Both can be wrong. Only one of them tells you who to ask.

What we do.

Soundmarque signs audio as a conformant C2PA product, listed on the Conformance Product List for both generating and validating audio. The recording leaves carrying a signed, machine-readable record of what it is and where it came from, and that record can be read with tooling we do not control. We are one of many conformant products, which is the point: a credential that only we can check would not be a credential.

What it is

An ISCC content code computed from the audio itself, so the identity comes from the recording rather than from a filename someone typed.

Who made it

The tool and the party that made the file, named in the record at the moment it was made, not attached to it afterwards.

What was used

Every ingredient that went in, with its relationship to the finished file, so a mix declares what it was built from.

When

An RFC 3161 timestamp from an independent authority, so the moment is provable without trusting us or anyone else.

Where the declaration comes from matters more than the flag.

A track-level AI flag is a single value with no author. A record instead says which tool made the file, when, and what it was built from, and keeps whoever declared it attached to the declaration. When a partner tool tells us what it did, that account travels in the record under their name, marked as theirs and not witnessed by us.

That distinction is dull until a claim is contested. Then it is the whole question: not whether the box was ticked, but who ticked it, when, and whether anything has changed since.

What this does not do.

A signature proves who said something, not that it was true.

When a tool declares that a model made a recording, the record carries that declaration, who made it, and the fact that it has not changed since. It does not make us a witness to the work. Where a declaration is somebody else's, the record says so on its face and names them.

A manifest can be removed.

Stripping it takes an ordinary transcode, not an attack, and anybody claiming otherwise is selling something. What a signature gives you is that the record cannot be quietly altered: a changed file stops verifying, and the failure is visible to whoever is checking.

None of it depends on us being around.

The record is checked with open tooling we do not control, against a certificate chain and a timestamp authority that are not ours. If Soundmarque disappeared tomorrow, every record we have signed would still verify. That is the only reason an attestation from us is worth anything.

Built on standards, not on us.

C2PA Content Credentials for the manifest. ISCC (ISO 24138) for the identifier computed from the audio. ISRC (ISO 3901) where the industry already has one. RFC 3161 for the timestamp, from a commercial authority with no stake in the record. Each of those is published, and each is implemented by people who are not us. A reader who distrusts every word on this page can still check the file.

Check one yourself.

A real record on a real recording, including its AI declaration. Nothing on it is drawn that is not a function of its own evidence.

Proving who said it — Soundmarque · Soundmarque