Audio — podcasts, radio, voice

When words are put in someone’s mouth, the original answers.

Veriframe verifies recordings the same way it verifies video and images: against the indexed original, never against a guess. Publish a speech, interview, or broadcast once — anyone holding a clip can check whether it still matches, and if it doesn’t, the verdict names the exact seconds where it diverges.

clip_from_social.mp31:14
0:41–0:44 · does not match the original
aligned_to_originalPosition 12:06 in the indexed recording
windows_compared23 of 24 match
verdictmodified — one window diverges

Why it matters

A spliced sentence has nowhere to hide

Compared window by window

The recording is checked in windows of a few seconds, and one divergent window is enough to break authentication. Averaging over the whole track would hide a replaced sentence; per-window comparison finds it — and names the seconds.

Clips trace back to their place

A thirty-second quote cut from an hour-long recording aligns back to its exact position in the original, even when the cut lands mid-sentence. The full context is one click away.

Everyday processing stays harmless

Re-encoding, resampling, and volume changes by platforms and podcast apps keep the recording anchored to its original — reported as re-encoded, never as tampering.

How VERA works for audio

The same registry, window by window

1

Index the original

Your team publishes the original recording to the Veriframe registry at release time. The registry stores its acoustic fingerprints in overlapping windows of a few seconds each — not just the file.

2

Anyone checks a copy

A journalist, fact-checker, or listener drops the clip they received into the public verify page. No account needed.

3

VERA compares window by window

The clip is aligned to the original by its loudness contour, then every few-second window is compared separately. One divergent window is enough to break authentication — and the verdict says which seconds diverged.

What the verdict tells you

Not "real or fake" — what happened to this recording

Every verdict states its evidence. These are the outcomes the audio pipeline distinguishes today:

authentic / re-encoded

Untouched or re-encoded

Platform recompression, format conversion, resampling, and volume changes keep the recording anchored to its original — reported as re-encoded, not as tampering.

re-encoded

Trimmed clips

A fragment cut from a longer recording is aligned back to its position in the original — even when the cut lands mid-sentence.

modified

Spliced or dubbed sections

A sentence replaced mid-recording lights up its exact window while the rest still matches. The verdict localizes the divergence to the second.

not found

Fully replaced audio

A recording that shares nothing with the indexed original is exactly that — different audio. VERA reports "not found" rather than speculating about what it might be.

Use cases

Built for the moment a quote is disputed

GOVERNMENT & PUBLIC OFFICE

Speeches and debates

Index the official recording when it is released. When a clip with rearranged words starts circulating, the verdict shows exactly which seconds diverge from what was actually said.

BROADCASTERS & PODCASTS

Episodes and interviews

Publish the episode and index it in the same step. Every quote clip in circulation can then be traced back to its place in the conversation — trimmed is fine, spliced is named.

FACT-CHECKERS

Voice clips under dispute

A contested clip takes moments to check against the registry, and the verdict — with the diverging seconds named — has a permalink you can cite.

Honest limits

The verdict stops where the evidence stops

We do not detect AI voices from a recording alone

Veriframe proves whether a recording matches an indexed original. It does not — and does not pretend to — listen to an arbitrary clip and rule whether the voice is cloned. That claim cannot be made reliably, so we never make it.

No original in the registry, no verdict

If the original was never indexed, VERA answers "not found" — honestly. The registry model is the point: proof, not probability.

Heavy background noise reads as difference

A copy buried under added noise may fail to match its original. The failure direction is deliberate: a degraded copy returns "not found" — never a false confirmation.

Try it on a real recording

The public verify page accepts MP3, WAV, M4A, OGG and FLAC alongside video and images.