VoiceGuard – Real-Time AI Voice Fraud Detection

€540M

APP fraud losses
UK alone, 2024*

+442%

Vishing incidents
H2 2024 vs. prior year

<150ms

Detection latency
real-time, on-call

97.4%

Accuracy
4-stream ensemble

The Challenge

📞

Voice fraud is the fastest-growing threat vector

Fraudsters call customers pretending to be bank employees — a technique called vishing. Increasingly, they augment these attacks with AI-generated deepfake voices, impersonating known contacts or executives.

Customer authorises payment under social engineering — bank is exposed
Deepfake voice cloning enables impersonation at scale, no human needed
UK PSR 2024: mandatory reimbursement up to €100K per case — banks now pay for the customer's mistake
PSD3 (~2026/27) will extend mandatory reimbursement across the EU

✅

VoiceGuard: detect, document, defend

VoiceGuard analyses every call in real time. It detects both human social engineering (vishing) and AI-generated deepfake voices — and builds a PSD2-compliant audit trail that supports a gross negligence defense in court.

Real-time warning delivered to the customer on the call
Documented evidence trail per transaction — dispute-ready
Bank builds a gross negligence defense, dramatically reducing reimbursement exposure
SDK integration — drops into existing call center infrastructure

Technology

4-Stream KAN Architecture — Interpretable AI

Stream 1

🌊

Spectral Analysis

Mel-spectrogram frequency patterns — detects unnatural harmonic artefacts in synthetic voices

Stream 2

🗣️

Prosody & Rhythm

Pitch, cadence, pause timing — catches AI-generated speech that lacks natural human micro-variation

Stream 3

🧠

Semantic Pressure

NLP-based detection of social engineering scripts — urgency, authority, fear-based manipulation patterns

Stream 4

📡

Channel Artifacts

VoIP, PSTN, codec fingerprinting — exposes audio path inconsistencies typical of deepfake injection

KAN Fusion Layer

Kolmogorov-Arnold Network

Unlike black-box neural networks (MLP/Transformer), KAN learns interpretable mathematical functions on each edge — every decision is auditable. No ReLU, no opaque weights. This is neurosymbolic AI: explainable to regulators, auditable for compliance teams.

⚡ <150ms

Target Markets

Three high-value segments, one SDK

🏦

Retail & Corporate Banking

Protect authorised push payment (APP) fraud flows. Build gross negligence audit trail per call. Integrate into IVR and call center platforms.

PSD2 / PSD3 PSR 2024 APP Fraud

📲

Telco & MNO

Offer VoiceGuard as a value-added service on mobile/VoIP plans. Network-level deployment — no end-user app. New recurring revenue stream per subscriber.

Network VAS SIM Swap Per-subscriber

🎧

Contact Centers & BPO

Real-time agent assist: flag suspicious callers before the agent acts. Reduce fraudulent transfer authorisations. Works with Genesys, Avaya, Five9, NICE.

Agent Assist Real-time API CCaaS

Business Case

ROI Snapshot — Mid-Size Bank (5% UK market share)

Item	Annual (€)
Current APP fraud exposure	€12.5M
Projected loss after VoiceGuard	€5.0M
Annual savings	€7.5M
Enterprise licence (annual)	€500K
Net ROI	15× return

* Based on UK Finance Annual Fraud Report 2025 (~€540M APP losses), PSR per-institution data, 5% market share estimate. EU figures vary by jurisdiction; PSD3 reimbursement rules pending.

Compliance Coverage

✓ PSD2 Art. 73 — generates documented evidence trail for gross negligence defense
✓ UK PSR 2024 — supports mandatory reimbursement exception process
✓ GDPR Art. 5 — no biometric data stored; audio processed on-stream, not retained
✓ EU AI Act — KAN interpretability supports High-Risk AI system explainability requirements
✓ EBA Guidelines — aligns with advanced authentication and fraud monitoring obligations

Integration & Deployment

📦

SDK — Drop-in Integration

REST API + WebSocket streaming. Works with any telephony platform. No new hardware, no new app for end users. Estimated integration: 2–4 weeks.

☁️

Deployment Options

Cloud (multi-region), private cloud, or on-premise — all supported. Data sovereignty guaranteed: audio never leaves the customer's environment if required.

🔌

Platform Compatibility

Genesys Cloud, Avaya, Cisco CUCM, Five9, NICE CXone, Amazon Connect, custom VoIP / SIP trunks. Output: real-time risk score + structured JSON audit log.

Request a live demo or PoC proposal

We can deploy a 30-day Proof of Concept on your test environment.
No commitment. Full technical & business case documentation provided.

👤 Németh Gyula — Founder & Lead AI Engineer

📧 gyula.nemeth@bluewaveai.co

🌐 bluewaveai.co

Request PoC →