Detect VALL-E AI voices.

Updated July 2026

Suspect a clip was made with VALL-E? Drop it in and get a citable verdict with the model named, in under half a second. Microsoft research model that clones a voice from about three seconds of audio using neural codec language modeling.

Check a clip free All generators

Synthetic detectedVALL-E

Confidence

high

Model

VALL-E

Yes, you can check. Drop the clip into the detector and it returns the probability that the speech was generated by VALL-E, a confidence level, and the named model, in under half a second. It reads the audio's synthesis signature, so it works on cloned and stock VALL-E voices alike, and gives you a citation URL to quote.

What it is

VALL-E, and why it shows up in suspicious audio

VALL-E is a Microsoft Research text-to-speech model that can reproduce a speaker's voice, and even the acoustic feel of the original recording, from roughly three seconds of sample audio, using a neural codec language-modeling approach. It is a research system rather than a consumer product.

Its approach is influential, and it illustrates how little audio cloning now needs. Most VALL-E clips in the wild are demos or derivatives, and the detector reads the synthesis signature regardless of the exact build.

Where you tend to see it: Research demos, derivative tools, and proof-of-concept impersonation.

VALL-E is a legitimate product; misuse is the problem, not the tool. You can read about it on the official VALL-E site.

The tells

How to tell a VALL-E voice

The human ear is unreliable on current VALL-E audio. These are the signals a detector weighs. We report which ones drove the verdict rather than handing you a bare number.

01Neural codec artifacts that a model reads even in short clips.
02Short-sample cloning that reproduces room tone in a slightly off way.
03Prosodic regularity across a passage.
04Transitions between phonemes that are cleaner than natural speech.

Spectral view · artifacts concentrate where synthesis smooths what a human voice would not

How the detector identifies VALL-E

A verdict you can cite, not a vibe

The detector reads the audio, not the speaker. Cloning or stock voice, clean or compressed, it looks for the synthesis signature and attributes the source.

Step 1

Drop the clip

Upload a file or paste a URL. MP3, WAV, M4A, WebM, or the audio track of a video. About half a second of clear speech is enough.

Step 2

The model scores it

The same model behind the public detector reads the acoustic signature and weighs the artifacts, then attributes the source, e.g. VALL-E, when it recognizes it.

Step 3

Get a citable verdict

You get a probability, a confidence level, the named model, and a permanent citation URL you can quote, file, or subpoena.

0.48s

Median verdict

99%

Accuracy on clean audio

24+

Generators covered

24h

Audio deleted after

If a clip turns out synthetic

What to do next

A verdict is evidence, not a verdict of intent. Save the result to get a permanent citation URL and a one-way audio fingerprint you can reference later without storing the file. If you are a journalist or investigator, cite the verdict alongside your own reporting; if this is a suspected scam, treat the contact as unverified and confirm through a channel you already trust. The FTC's advice on suspected scam calls is a sensible baseline.

Building this into a workflow? The API returns the same verdict as JSON with webhooks for bulk jobs, and the browser extension checks audio in place on WhatsApp Web, YouTube, and podcasts.

FAQ · VALL-E

Common questions

Can you detect VALL-E output?

Yes. The detector reads the synthesis signature the codec-modeling approach leaves behind, with confidence depending on clip length and quality.

VALL-E is research, so does this matter?

The approach has spread into derivative tools, and the detector keys on the synthesis traces rather than a specific brand build, so related outputs are in scope.

Does a three-second clone evade it?

Short cloning does not remove the signature, though very short clips lower confidence, which we report rather than hide.

Is a single check free?

Yes, a single verdict is free with no card.

Will it name VALL-E specifically?

When recognizable we name the source; otherwise we return 'unknown synthesis'.

Is this a VALL-E voice? Find out.

Free verdict, model named, in under a second. No card to start.

Open detector Use the API