Upload an audio clip and get a verdict in seconds — whether the voice is a real human or synthetic speech from tools like ElevenLabs, OpenAI, Murf, PlayHT or Resemble.
or click to browse — MP3, WAV, M4A, OGG, FLAC up to 10 MB and 2 minutes. Your audio is sent over an encrypted connection, processed, and discarded — it is never stored. Privacy policy.
Drop an audio clip or try a sample. You'll get a verdict, confidence score and file details.
Every plan below is for the AI Voice Detector only. A TextSight text subscription (Starter, Pro or Business) does not include voice checks, and a voice plan does not include the text tools.
Your audio is analysed on our own GPU servers — it never goes to a third-party model provider and it is never stored. For banks, KYC vendors and call centres, that is usually the part that decides the deal.
Clip length is the longest file a plan accepts; the detector currently scores the first 60 seconds of each clip. Yearly billing is charged once for the twelve months and saves about 30% against paying monthly, and code VOICE50 takes another 50% off your first year on Echo, Amplify and Broadcast, which works out to about 65% off the monthly rate for that year. Manage or cancel any time from your account — see full voice pricing.
AI voice tools — ElevenLabs, OpenAI, Murf, PlayHT, Resemble — now clone and synthesize speech convincingly enough to fool the ear. Journalists worry about fabricated quotes, fraud teams about voice scams, and creators about cloned voices.
TextSight's AI Voice Detector analyzes spectral features, micro-pauses, breathing and waveform characteristics that synthetic voices leave behind. Upload a clip, get a verdict in seconds.
Synthetic speech can sound natural to the ear, but it is produced very differently from a human voice — and that difference shows up in the audio signal. Text-to-speech and voice-cloning models reconstruct a waveform from a learned model of speech, which tends to smooth over the tiny irregularities a real microphone always captures.
The detector inspects spectral features across the frequency range, the rhythm of micro-pauses between words, breathing and mouth noise, and the way energy rises and falls inside each syllable. Human recordings carry inconsistent room noise, natural breaths and slight timing variation; many AI voices are cleaner and more uniform than a real recording ever is. The tool weighs these signals together and returns a verdict with a confidence score, rather than a plain yes-or-no.
AI voice cloning has moved from novelty to everyday risk — a few seconds of someone's voice can be enough to generate convincing fake speech. That is why more people now need a fast way to verify audio:
The detector looks for the traces synthetic speech leaves in the signal rather than matching a vendor, so it is aimed at text-to-speech and voice-cloning engines in general — ElevenLabs, OpenAI, Murf, PlayHT, Resemble, Amazon Polly, Azure and Google voices among them. We have not published a per-engine evaluation, so we do not claim a tested detection rate for any named tool, and new engines appear faster than any detector is retrained.ElevenLabs, OpenAI TTS, Murf, PlayHT, Resemble, Amazon Polly, Microsoft Azure TTS and Google Cloud TTS — plus most consumer AI voice apps built on top of them.
Because new voice models appear constantly, no detector can promise to catch every one. Heavily compressed, noisy or re-recorded audio lowers confidence for any tool — so treat the score as evidence to weigh, not absolute proof. We explain exactly what moves reliability up and down, and why we publish no accuracy percentage, on our accuracy page.
Run a check in seconds — first one free, no signup, no audio stored.
Analyze audio