It has his voice, his phrasing, his impatience. It asks for $48,900 to a new account before 5:00, and it asks you to keep accounts payable out of it. Someone in finance plays messages like this one every week, and the wires that leave are real.
Nothing here is a feeling. The voice does not match the enrolled voiceprint, the caller ID does not match the directory, the word "new" arrives 31% too fast, and the secrecy line is the phrase a cloned-voice fraud almost always reaches for. Each one is a panel on its own, starting now.
The voicemail arrived from the main office line, the one in the lobby directory. That sounds right. It is exactly what you would expect a CFO calling from the road to dodge.
The switchboard number is spoofed: the call never traversed your PBX. The carrier handshake on the recording shows a SIP trail from a retail voice provider two states away, which your phone system logs and Earprint reads.
A clone is trained on recordings: a webinar, an earnings call, thirty seconds of a conference talk. It copies the voice. It cannot copy the way a real voice drifts, the small wobble a throat makes on a long vowel, the breath it takes mid-sentence.
Steady. Every vowel holds its pitch. No breath before "wire", no swallow on "Halston". A performance with nothing underneath it.
Coughs at 0:11, takes two breaths in one sentence, trails off on the last word. A throat, in other words.
Earprint keeps a voiceprint built from 41 genuine voicemails Greg left this year. Tonight's caller is plotted against that cluster. It does not land in it. It does not land near it.
Distance from cluster: 4.7σ. Anything past 2.5σ fails.
You can hear it if you know to listen: "the new account is" comes out a shade too quick, on a breath that doesn't belong to the sentence. That is where the message was assembled. The bank details were never in the original call at all.
"...to the Halston account. They moved banks, so it's the new account is details I texted you..." · the grammar stumbles exactly where two recordings meet. Nobody notices at 8:00 in the morning on a speakerphone in a lobby.
At 23.1 seconds the speaking rate jumps and the room tone changes. The words "new account" run 31% faster than the man's average, on a different breath profile. The waveform has a seam.
Room tone before the seam: your CFO's home office, measured on Tuesday's genuine call. After the seam: a treated room that matches no space he has ever called from.
"Before five" and "don't loop in accounts payable." A real CFO with a real deadline might ask for speed. No CFO asks you to hide the payment from the people whose whole job is paying it. If a voice message ever asks you to keep a payment quiet, that alone is the reason to stop.
Speed, a deadline inside fifteen minutes. Authority, the boss's voice and the boss's impatience. Secrecy, the one person who would catch it is told to stay out. Finance teams that train on those three words catch most of these with the phone still in their hand.
Earprint reads the transcript for the pattern, because the pattern is more reliable than any voice: speed, authority, secrecy, and a payment channel change. Three of the four together score higher than any single technical signal.
The same man, the same phone, the same kinds of words about money. Every signal checks, and the check says so in plain numbers, because a check that can only say no is not worth running.
"It's Greg. Auditors want the Q3 freight ledger before Thursday, nothing urgent tonight, just have Alicia pull it in the morning. I coughed, sorry, cold is still hanging on."
"It's Greg. I need a wire out today, forty-eight nine, to the new Halston account. Don't loop in accounts payable. Before five."
A verification tool that only ever screams is a tool nobody trusts by March. Earprint returns the same four signals on every message, and the numbers go in your incident log untouched: the sigma distance, the carrier trail, the splice second, the transcript pattern. Your controller re-verifies by hand, and the log is the record the auditor and the insurer both ask for.
Green, with the four numbers beside it. Nobody acts on a pass. On a fail, Earprint texts the enrolled executive's real mobile and holds the message in your queue until a named human clears it, and the hold is on the record too.
It sits beside your PBX or your VoIP carrier, hears every voicemail and inbound call to a finance seat, and returns a verdict before the transcription finishes. Enrollment takes thirty seconds of real audio per voice, which for most teams is the last all-hands recording.
Finance and accounting seats at companies from 5 to 2,000 people, and anyone whose boss can ask for a payment by phone: property managers, school business offices, law firms.
Both, on the numbers you point it at. Live calls get a verdict in about two seconds and can be flagged to a second approver before the wire window closes.
Audio never trains anyone else's model. The voiceprints are yours, and the retention window is yours, down to zero.
Earprint checks voices. It does not check texts, emails, or a caller who says nothing. Here is the whole list.
A request that arrives only as text, an email from a lookalike domain, an invoice that "just has new bank details": those never reach a microphone. Pair Earprint with a call-back rule, not instead of one.
A heavy cold, a new headset, a call from a moving car: all of it moves the numbers. Real changes pass after a thirty-second re-enrollment, and the check tells you when it is unsure rather than guessing.
94% confidence is a reason to stop and call back on the enrolled number. It is not a legal finding, and the hold it puts on a message is released by a named person, always.
"The Halston message was the third clone aimed at us this year and the first one that never had a chance. The controller had the verdict on her screen before the message finished playing."
ALICIA SANDOVAL, CONTROLLER, BLUEBONNET FREIGHT CO. · FORT WORTH, TEXAS
Message, verdict, the four signals with their numbers, who cleared or held it and when. Retained per your policy, exportable as the incident log your insurer's cyber form asks about. On a fail, the hold and its release are on the same line.
SIP, Teams, or a carrier forward. Most teams are hearing checks the same afternoon.
Thirty seconds of real audio each: the CFO, the controller, anyone who can ask for a wire.
Every voicemail to a finance seat gets the four signals and a verdict. The first clone attempt is usually inside the month.
The same three steps on either side of the switch. Enrollment is thirty seconds of real audio per voice, and the trial runs thirty days without a card.
Every message checked, every signal scored, including the ones that passed. It is the baseline your controller signs.
On a fail, a named human clears or holds the message and the record keeps both. The check is evidence; the decision is always yours.