AI Voice Detector

Yes. Free, online, no sign-up, no upload limit and no watermark on the result. It runs entirely in your browser, so there is no server cost to pass on to you.

No, and be suspicious of any tool that says it can. Even paid detectors trained on large datasets publish error rates, and they slip further every time a new voice model ships. Treat any detector, this one included, as one input next to context: who sent the file, what they want, and whether the story checks out elsewhere.

Accurate enough to be useful, not accurate enough to settle an argument. The high percentages vendors advertise are usually measured on audio the vendor picked, and independent testing in the field has found that error rates rise sharply once a clip has been compressed, sent through a phone network or produced by a generator the detector was never trained on. A real phone recording is close to the worst case for every tool on the market.

It is not tied to any one generator. It looks at properties of the audio itself, so it reacts to whatever a model leaves behind rather than to a vendor fingerprint. In practice older or lower-tier voices from any provider show more of these signals than a current flagship voice does. ElevenLabs also runs its own classifier for audio made with its models, which covers that one generator in more detail than any general checker can.

Whatever your browser can decode: MP3, WAV, M4A, OGG and usually FLAC. If a file will not open, convert the clip to WAV and try again. Longer is better, since pacing and breathing need something to measure. Under six seconds of speech and the tool will say it cannot judge.

Extract the audio first, then upload the file. Be aware that platform audio has been re-encoded at least once, which eats the high frequencies and can push a real human clip toward an AI-looking reading. Weigh the pacing and breathing signals more heavily than the frequency one, and check whether the channel used a different voice in older uploads.

Several browser extensions score audio as it plays in a tab, which suits YouTube and video calls better than a file uploader does. Before installing one, check whether it analyses on your own device or sends the audio to a server, since that decides whether a private recording stays private. This page itself works on a phone browser, so there is nothing to install for a file you already have.

Yes, the research code is public. Architectures such as AASIST, trained and scored on the ASVspoof datasets, are what most academic and commercial claims trace back to, and you can run them yourself with Python. Expect research-grade setup rather than a click-and-go tool, and pay attention to results measured on attacks the model never saw in training, because those are the honest numbers.

No. Those are separate problems, and doing them in a browser badly would only produce confident nonsense. Accent and gender classifiers describe a speaker rather than a file, so they will happily classify a cloned voice too. This tool answers one question: does the audio carry the signs of being generated rather than recorded.

Far less than people expect. A few seconds of clear speech is enough for current cloning tools, which is roughly the length of a voicemail greeting or one sentence from a posted video. That is why a family code word and a callback on a number you already had are worth more than any listening test.

Hang up, then call the person back on the number already in your contacts, not one the caller gave you. Do not send money, gift cards, crypto or a one-time passcode while the call is live, and do not stay on the line to test them, since every second of your voice is material. If you cannot reach them, call someone else who would know where they are, then report it to your bank and your country’s fraud line.

Never. There is no upload step. Open your browser network tab and watch if you like: the file is read straight into memory on your machine and discarded when you leave.