Listening closely is not a reliable test any more: current voice generators produce breaths, pauses and room tone on purpose. What works is a short routine — get the cleanest copy you can, run it through a detector, read the result for what it is, and settle the question through something other than the audio. This page walks through it, including the parts no detector can do for you.
Before you analyse anything, ask what you would do with the answer. If a recording is asking you to send money, share a code, or act urgently, the fastest and most reliable check is not acoustic at all: hang up, and call the person back on a number you already had or look up yourself. That settles who you are talking to in a way no detector can.
A detector is the right tool when the recording itself is the evidence — a clip you are about to publish, a voice note someone forwarded, a narration you were sent as finished work, an interview answer submitted as a file. In those cases, the steps below give you a result you can defend.
Each one removes a way the result could mislead you.
.opus note, an iPhone audio message, a video — convert it a single time, preferably to WAV so no further compression is added. The voice message guide and video voiceover guide cover the common cases.Every check returns a verdict and an AI-likelihood score between 0 and 100. Read them together, and read them against the conditions of the file.
| You see | What it suggests | What to do next |
|---|---|---|
| Likely an AI voice | The clip carries features the model associates with synthetic speech. | Treat it as a reason to investigate, not a conclusion. Run a reference clip; verify the speaker by another channel. |
| Likely a real human voice | No strong synthesis evidence in the part that was scored. | This does not clear the recording of being edited, misattributed or taken out of context — those need other checks. |
| Inconclusive | The score sits in the uncertain middle, or the audio was too short, quiet or noisy to judge. | Find a longer or cleaner copy, or move to verification that does not depend on the audio. |
On paid plans in the voice app, results also include a timeline showing which stretches of the scored audio leaned synthetic, and a PDF report you can keep with your notes.
Sometimes, with older or cheaper voices: flat emotion that never shifts, odd stress on names and numbers, breaths that land in the wrong places, or a voice that sounds identical in every sentence regardless of mood. These are worth noticing, but they are not reliable. Modern generators add breathing and hesitation deliberately, and real people on bad phone lines sound strange too.
Treat your ear as a reason to check, not as the check. And treat a confident “it sounds fine to me” with the same caution, because convincing synthetic speech is exactly the scenario detectors exist for.
Use a known-real clip of the same speaker as a baseline.
Read the method →Getting WhatsApp, Telegram and iPhone notes into a checkable file.
Read the guide →What moves a result up or down, and what false positives look like.
See our position →Your first check is free with no signup — 3 a day with a free account. Your audio is processed and discarded, never stored.