Audio Deepfake Detector

How this audio deepfake detector works

An audio deepfake detector reviews a saved voice note, voicemail, or call recording for cloned or synthetic speech. This page is that check. Upload an MP3, WAV, M4A, AAC, OGG, Opus, FLAC, WebM, or AMR file up to 50 MB. The service sends that speech to Google Gemini with forensic guidance. It does not run a separate laboratory spectrogram, speaker-authentication model, or vendor classifier. It does not prove who is speaking. Analyzed content is retained for up to 7 days, then deleted. A limited free allowance applies. No signup.

If the file is a picture, use the AI image detector. If it is a clip, use the deepfake video detector or the face swap detector. The general upload is the deepfake detector. Related reading: AI voice family safety. The same service is the French détecteur vidéo IA and the Russian detector; French and Russian audio pages are /fr/audio/ and /ru/audio/.

The review listens for audible clues, and only when more than one of them lines up:

• Breathing and mouth sounds. Missing or repeated pauses can be a clue. Noise removal can create the same effect, so this is never enough on its own.
• Prosody. Unusually flat or uniform delivery can matter when other speech in the same clip sounds different.
• Noise continuity. Sudden digital silence or a metallic edge on sibilants can warrant a closer look. Phone and messaging codecs can imitate both.

What an AI voice detector can and cannot detect

Use the result to decide whether a recording deserves a second look. Do not use it as the decision itself.

• Can review: a saved voice note, voicemail, or call recording for patterns associated with cloned or synthetic speech.
• Cannot prove identity. A familiar voice can be cloned, and a real person can sound strange on a bad connection. The score is not proof of who spoke.
• Cannot name a product. The check does not attribute audio to a specific voice-clone vendor, and it does not verify invisible watermarks or vendor labels.
• Cannot listen to a live call. Analyze a recording you upload. While a call is still in progress, hang up and call back on a number you already trust.
• Not for music. Singing and instrumental tracks are outside this speech review.
• No guaranteed verdict. There is no single accuracy figure that holds for every voice, codec, and edit. Short, noisy, or denoised clips should be read as lower confidence.

How to check a voice message or call recording

1. Save the voice message, voicemail, or call recording as a file. The original export is more useful than a phone held up to a speaker.
2. If only one sentence sounds wrong, trim around that sentence and keep a little of the speech before and after it.
3. Upload the file here. Files up to 5 MB are read as provided. A larger file is converted in the browser to one representative segment of up to 30 seconds at 16 kHz mono.
4. Read the result as a screening signal. For a request involving money, access, or identity, contact the person through a channel you already know, even if the score looks low.

Limits of voice-clone screening

Very short clips contain less evidence. Telephone codecs, platform compression, and aggressive noise removal can erase useful detail or imitate synthetic artifacts. If the suspicious phrase sits outside the 30-second segment used for a large file, the check can miss it until you trim and run that part again. A low-risk result does not mean the speaker is real. A high-risk result does not prove fraud or name the tool that made the voice.

When people use this check

• Families triaging an urgent voice note before sending money.
• Newsrooms and podcasters reviewing a clip of a public figure before quoting it.
• Finance and support teams screening a recorded impersonation before a payment or account change.
• Voice actors and creators taking a first look at a clip that may use their voice without permission. Specialist review still belongs on any formal dispute.