Upload a call or voice-message recording; Whisper transcribes it inside your browser and checks for common voice-phishing phrases. The audio never leaves this device.
Free with no sign-up, and the check runs inside your browser, so what you enter never leaves your device. The result is a signal to help you decide, not proof for money or legal decisions.
Upload a call recording and it is transcribed at the edge (the audio is never stored), then the text is scanned with the same rules as the smishing check to flag phrasing typical of voice-phishing. It looks at what was said — not at whether the voice is real.
Two scripts cover most calls. The authority script opens with "an account in your name is linked to a crime", moves through prosecutors, regulators and police, and ends at "move your money to a safe account and we will return it after the investigation". The loan script dangles a low-rate refinance and demands an upfront payment to "clear the existing loan first". Both end at a request no real institution makes.
The wording changes; the skeleton does not — authority (agency, title, case number), urgency (today only, do not hang up) and an abnormal payment route (safe account, cash withdrawal, gift cards, crypto). When all three show up in one call, you are listening to a script. The result panel breaks its findings down along the same skeleton.
A few seconds of public video is enough to clone a voice. It is the "Mum, I lost my phone" message scam moved to audio, and the problem is that human ears are poor at catching it.
So the defence is procedure, not detection. Hang up and call back on the number already saved in your phone — never the one you were given. Use a check question your family agreed on in advance; a cloned voice cannot clone memories. Send nothing until that call back succeeds.
If the transcript is wrong, the verdict is wrong. Accents, background noise and technical terms are misheard often, and a bad line can garble whole sentences. Read the transcript on screen first; if it does not match the call, ignore the result.
This also looks for known phrasing. A freshly written script, or an approach disguised as small talk, can score low. It does not judge whether the voice itself was synthesised. That is outside this tool.
Checking what the caller claims is genuinely simple: hang up and call back on the official number you already had. Not the number they gave you, not the one on your screen — the one you held before this call started.
For anything financial the channels are fixed. Your bank's number is on the back of the card or on its website; national financial fraud helplines exist in most countries; payment stops go through your bank and the police. Official lookup services that show which accounts and telecom services exist in your name are the fastest route when identity theft is the worry.
One caution. On a device carrying a malicious app, that call itself can be intercepted. If you installed anything or tapped a link during the call, make the verification call from someone else's phone.
The caller gave a case number, said an investigator needed cooperation because the listener's identity had been stolen, and warned that hanging up would trigger a summons. The call ended at a request to transfer funds to a "safe account" for inspection. The transcript lights up all three markers — impersonated authority, urgency pressure, abnormal payment route — and the risk reads high. The deciding factor is not the score but that final request: no law-enforcement body asks for a transfer over the phone, under any pretext.
Using a bank name, the caller said a low-rate switch had been approved, but the existing loan had to be cleared first before the limit could be set — with a partial payment to an account they provided. The call also pushed an app install. The bait (approved low rate) and the upfront payment demand are flagged together. No real lender collects repayment into a personal account before disbursing a loan, and the suggested app is often malware that intercepts outgoing calls.
A card issuer called first about suspected fraudulent use: identity verification, references to the account and recent transactions, and a "we need to handle this now". Institution names, account talk and urgency overlap with scam vocabulary, so the risk can rise. The check reads phrasing; it cannot verify the caller. The dividing line stays the same — hang up and call the official number on the back of the card. A legitimate case simply continues from there.
No. It is used for transcription and dropped immediately — neither the audio nor the transcript is saved. The result lives only in your browser and disappears on reload.
It depends on where you are. In Korea, recording a call you are part of is permitted (the law targets interception by a third party). In the United States some states require all-party consent. Check your jurisdiction before you rely on a recording.
No. It means no known phrasing was found. If you were asked for money or personal data, follow the hang-up-and-call-back procedure regardless of the score.
Call your bank and the police immediately to request a payment stop. If the funds have not moved on to the next account, they can often be frozen — speed is everything. Keep the recording and the transfer records.
The smishing check is the right place for text. This page only adds a speech-to-text step in front; the rules applied to the sentences are the same. If it is already written, skipping transcription is more accurate.
Transcription runs on this device — the audio is never uploaded. The speech engine (~76MB) downloads once. Max 8MB.
Transcription and pattern checks are reference only. For agency impersonation or money requests, verify via the official number regardless of the score.