Audio authentication.
When a recording is disputed, you need more than a hunch. Get a citable, reproducible verdict on whether a voice is real or AI-generated, with the source model named, in under half a second.
Authentication, in the practical sense
The word "authentication" gets used loosely, so it is worth being precise. This is not verifying who is speaking, and it is not a watermark. It is answering one question that increasingly decides disputes: is this voice real, or was it synthesized? Because that answer is now contested in fraud cases, custody matters, employment disputes, and newsroom decisions, it has to be more than an opinion. It has to be repeatable and citable.
The audio, measured
The detector reads the acoustic fingerprint of the clip, the spectral, prosodic, and timing traces synthesis leaves behind, rather than the words. It works on cloned and stock voices and names the generator where it recognizes one, returning an honest "unknown synthesis" when it cannot attribute the source. The method is published and version-stamped, so the same test can be reproduced later against the same snapshot. See how to verify AI audio for the step-by-step.
A record, not just a number
How likely, honestly
A 0 to 1 likelihood the speech is machine-generated, never rounded up, paired with a confidence level driven by clip length and quality.
The source, named
The recognized generator where the signature is clear, or a plain "unknown synthesis" when it is not.
Reproducible record
A permanent URL, a methodology version stamp, and a one-way fingerprint of the file, so an opposing expert can rerun the test.
Disputes, investigations, and the newsroom
Investigators and legal teams authenticate recordings that enter disputes and cases; see the legal and investigations use case. Fraud teams authenticate flagged voicemails and call recordings before acting; see fraud detection. Newsrooms authenticate audio before publishing; see newsrooms. For the underlying detection, see deepfake voice detection, and to build it into a workflow, the API returns the same verdict as JSON.
Analysis, not a ruling
We authenticate the audio, not the speaker, and we analyze recordings after the fact rather than monitoring live calls. We provide a probability with a published method behind it; we do not determine court admissibility, we are not a law firm, and we never claim a guaranteed 100% result. We are voice and audio only, so we read a video's audio track but not its images. For provenance on genuine media, content-authenticity standards like C2PA attach tamper-evident origin data and complement detection.
Common questions
What is audio authentication?
Is this the same as verifying a speaker's identity?
Can an authenticated verdict be used as evidence?
How accurate is it, and what are the limits?
Can we authenticate audio at volume?
Authenticate a recording now.
Free for a single verdict. API and volume on Starter and above.