Voice deepfake detector
Heuristic check for whether a vocal recording was likely produced by a human or by a TTS / voice-cloning system. Looks at spectral flatness, pitch jitter, prosody variance, and high-frequency energy. This is a weak signal, not a verdict — use the per-feature breakdown to judge for yourself.
Questions answered
Frequently asked questions
No — and it isn't close. What you get is a suspicion score, deliberately not called a probability, plus a verdict band. A quiet, closely-mic'd podcast recording can score much like a vocoder, and a heavily processed real vocal often scores high. Treat it as one weak signal among several. It must never be the sole basis for a ban, a takedown, or an accusation.
It measures four things that the anti-spoofing literature calls weakly discriminative: spectral flatness, cycle-to-cycle pitch jitter, how much the spectral centroid moves (a proxy for prosody), and how much energy sits above roughly 7 kHz, where many open-source vocoders truncate. Those four are combined into a single score, and every sub-score is shown so you can see which one moved the needle. Unless an operator has installed a trained anti-spoofing model, that heuristic is the whole detector.
MP3, WAV, FLAC, OGG, M4A, or WebM, up to 30 MB. It's free — no credits are deducted — but you do need to be signed in, and checks are rate-limited to 20 every five minutes.
Short clips, low bitrates, heavy compression or noise reduction, background music under the voice, and the newest synthesis models, which are specifically good at reproducing the micro-variation this heuristic looks for. When in doubt, read the per-feature breakdown rather than the headline number.