Report Bug / Feature Request

AI Voice Deepfake Detector

Spot AI-generated, cloned, and synthetic speech in any audio clip. Get a probability score plus confidence rating in seconds — the inverse of TTS, built on the same models.

Upload Audio to Verify

200 characters

Drag & drop your file here, or browse

Max 50MB per file. First 60 seconds are scored, long clips are truncated.

— or record from your microphone —

00:00

Free — 10 detections per day, no signup needed

How it works

The detector is a HuBERT-base classifier (Meta AI, MIT licensed) fine-tuned on FoR, ASVspoof, and WaveFake — three of the largest public corpora of human-vs-synthetic speech.

It returns the probability that the clip was generated by a TTS model, voice clone, or other AI system, plus a confidence rating. Treat the score as a strong signal, not legal proof.

Who uses this

Journalists vetting leaked or anonymous recordings
Marketplace buyers verifying voice samples are human before paying
Trust-and-safety teams flagging voice-cloned harassment
Researchers labelling datasets at scale
Creators proving their own audio is genuine

Supported Formats

MP3, WAV, FLAC, OGG, M4A

Max 50MB per file. First 60 seconds are scored, long clips are truncated.

Deepfake Detection Plans

Start free, upgrade when you need more

Free

10 detections per day
Clips up to 60 seconds
Probability + confidence score
No account required

Frequently Asked Questions

It analyses an audio clip and tells you the probability that the speech is AI-generated, cloned, or otherwise synthetic. You get a 0-100% score plus a confidence rating, so you can decide whether the recording is trustworthy enough for publication, payment, or evidence.

A HuBERT-base classifier (Meta AI, MIT licensed) fine-tuned on three of the largest public corpora of human-vs-synthetic speech: FoR (Fake-or-Real), ASVspoof, and WaveFake. The fine-tune is published as MelodyMachine/Deepfake-audio-detection-V2 on Hugging Face.

On the public FoR test set the underlying model reports above 95% accuracy, and on ASVspoof-style attacks it consistently flags TTS and voice-cloning output. Accuracy on novel models the classifier has never seen is lower — treat the score as a strong signal, not legal proof.

You can upload clips up to 50MB and roughly 60 seconds of speech. Longer recordings are automatically truncated to the first minute, which is enough for the classifier to make a stable decision.

No. The clip is processed in-memory on our GPU server, the result is returned to you, and the upload directory is deleted immediately. We never log, share, or use your audio for training.

Detection costs 200 characters per clip — the smallest paid operation on the platform. Free accounts can run a few detections per day; signed-in paid users can run them at full speed with priority queue access.

Trust matters more, not less, in a world full of synthetic voices. The marketplace lets buyers verify a creator-uploaded sample is human before paying. Journalists can vet leaked-recording authenticity. Platform safety teams can spot voice-cloned harassment. The detector is the inverse of TTS: same expertise, opposite goal.

Not yet — the v1 detector only outputs real vs synthetic. Model attribution (which TTS engine made this?) is on the roadmap once we have enough labelled samples per model.

5.0/5 (1)

AI Voice Deepfake Detector

Upload Audio to Verify

How it works

Who uses this

Supported Formats

Deepfake Detection Plans

Frequently Asked Questions

What does the deepfake detector do?

Which model powers the detector?

How accurate is the result?

How long can my audio be?

Is my audio stored after detection?

How much does it cost?

Why does this exist on a TTS site?

Can it tell which TTS model produced an audio clip?

Ready to get started?