Your bank's voice channel can be attacked 21 ways.
Most fraud teams defend against twelve.

VoiceGuard is a real-time deepfake & vishing detection layer for the bank voice channel — passive on the call (SIPREC), explainable in the verdict, and evidence-grade in the audit trail.

<1%
false-alarm rate
(session-level FPR)
>90%
recall on
held-out benchmarks
2.62%
system EER
(ASVspoof 5)
~36–40 ms
per-chunk latency
(production hot path)
The threat surface

21 vectors. 21 documented cases.

Every vector below maps to a real, published case or a regulator-recorded trend — not a hypothetical. This is the fraud map of the bank voice channel.

A · Against the customer

7 vectors
A1Bank / authority impersonation (vishing)
A2Authorised push payment (APP) fraud
A3Customer calls the bank on a scammer's script
A4Cloned-relative "grandparent" scam
A5SIM-swap → account takeover
A61:1 deepfake voice call
A7Multi-party deepfake conference (Arup, $25M · CNN 2024)

B · Against the bank

8 vectors
B1Voice-clone → IVR Voice ID bypass
B2Voice-clone → internal / treasury fraud
B3Human social engineering vs the agent
B4SIM-swap → OTP bypass
B5Insider / rogue agent
B6Synthetic-identity / voice-KYC bypass
B7Account-recovery hotline engineering
B8IVR reconnaissance / brute-force

C · Infrastructure

3 vectors
C1Caller-ID spoofing (even the regulator's number)
C2Call-forwarding exploit (*21# trick)
C3OTT channel-hopping (WhatsApp / Signal)

D · Asynchronous

3 vectors
D1Smishing → vishing escalation
D2Recovery-room scam / re-victimisation
D3Async voicemail / voice-note deepfake

VoiceGuard maps the full surface and defends across it in layers — passive on the call for the bank-side vectors, with an honest position on the rest (below).

What VoiceGuard does

One pass. Two threats. A single verdict.

Pillar 1 · Deepfake

Is it a live human voice, or AI-generated?

A 4-stream KAN ensemble scores every inbound caller automatically — no enrolment, no voiceprint. It closes the gap voice biometrics can't: a perfect clone passes biometric "match", and VoiceGuard asks the orthogonal question — live or synthetic?

Pillar 2 · Vishing / SE

Is a manipulation pattern being spoken?

Real-time speech recognition plus a fraud-keyword engine runs on both legs — caller and agent — and fires even without a deepfake: safe-account pretext, OTP coaching, channel-hopping, or an agent policy violation. Deterministic, immediate supervisor alert.

Dual-threat verdict. The system natively distinguishes VISHING_SCAM, DEEPFAKE_DETECTED and DEEPFAKE_VISHING — each pillar signals on its own and together.
How it connects

Passive on the line. Zero added latency.

Two tap points

A passive SIPREC media fork at the contact-centre SBC (inbound, IVR, agent) and the internal PBX — Cisco, Avaya, Webex or MS Teams compliance recording.

No risk to the call

Nothing sits in the live audio path: zero added latency, zero availability risk — DORA-friendly. It reuses your existing MiFID recording tap.

Decision-support, not control

VoiceGuard delivers structured risk telemetry to your fraud / SIEM / CRM. Your own systems and policies decide and act. An in-app, customer-activated check is one additional layer — not the headline.

Why it's different
01 · GLASS-BOX

Explainable verdict

A KAN architecture yields a verdict you can defend — per-model contribution and the rule chain that triggered it, not a black-box score.

02 · PASSIVE

No agent load

No caller enrolment, no friction in the conversation. It runs silently alongside the agent.

03 · EVIDENCE

Audit-grade trail

Every signal is timestamped, structured forensic evidence — the documented-diligence record a dispute or regulator asks for.

04 · NON-COMPETITIVE

De-risks your stack

It doesn't replace your Voice ID, fraud engine or CRM — it adds the orthogonal "live or AI?" signal that de-risks them.

Why now

With mandatory APP reimbursement (UK), fraud liability transfer (UAE) and DORA / PSR raising the bar, an undetected synthetic-voice transfer increasingly lands on the bank — and "we did nothing" becomes a documented decision in a dispute. The voice channel is where the cost is now concentrating.

Honest scope

In scope, in production today: synthetic-voice detection and vishing-pattern detection on calls reaching your contact-centre and internal PBX (passive SIPREC). Out of scope, stated plainly: video deepfakes, end-to-end-encrypted audio after a mid-call switch to WhatsApp/Signal, and pure DTMF IVR reconnaissance (a bank-side telephony control). Some customer-side vectors depend on mobile-operator (CAMARA) availability. We'd rather tell you the boundary than oversell past it.

The revenue angle — beyond protection

VoiceGuard doesn't have to be only a cost centre. A bank can offer it as an opt-in customer subscriptionyour bank watches and protects you, and the people you care about — turning fraud defence into a new revenue line, not just loss reduction. The model is already proven in adjacent categories (Carefull, sold through 35+ financial institutions; EverSafe; Fifth Third Identity Alert), and consumer willingness to pay for scam protection is well-documented. A bank-branded deepfake-voice subscription doesn't exist yet — that's the opening. It's usually a second conversation, with your retail or product team.

See it on your own audio — in a week, for free.

The Audio Challenge runs a structured detection test on your own (anonymised) samples and sends back the report. No integration, no personal data, no commitment. Or start with a 15-minute call.

Start an Audio Challenge Book a 15-minute call