Solution
Synthetic voice detection for banking and wire-fraud channels
Sonotheia detects AI-generated and converted voice in banking and FINRA-supervised workflows (wire authorizations, callback verification, and high-value instructions) using physics-based signal analysis validated on ASVspoof5 benchmarks. Every alert includes an explainable decision trace for fraud ops and financial services compliance review.
Wire fraud driven by synthetic voice is a biometric circumvention threat. Detection without governance leaves fraud teams with a score and no defensible record when the transfer already moved.
Which banking voice channels are highest risk?
Year-end and quarter-close periods concentrate risk: time-sensitive wire requests, vacation coverage callbacks, credential resets, trading instructions, and exception approvals where staff recognize a familiar voice and override policy. Synthetic media attacks target these operational weak points.
How does synthetic voice detection differ from keyword monitoring?
Keyword and phrase rules cannot detect novel synthesis pipelines. Sonotheia analyzes acoustic physics (spectral envelope dynamics and excitation patterns) that persist across TTS engines, voice conversion, and replay attacks after codec compression.
How does Sonotheia integrate with existing fraud stacks?
API-first integration with case-management exports. Sonotheia complements existing voice biometrics and transaction-monitoring tools by adding explainable spoof detection and governance documentation rather than replacing core banking systems.
Frequently asked questions
- How does Sonotheia support FINRA expectations for financial services voice fraud?
- Explainable decision traces, calibration version stamps, and vendor-oversight documentation help broker-dealers and banks map synthetic voice controls to FINRA supervisory and model-risk expectations without relying on opaque scores alone.
- Can Sonotheia detect ElevenLabs or other commercial TTS clones?
- Calibration includes contemporary TTS and voice-conversion attack families from ASVspoof5 and internal self-generated corpora. Specific engine coverage depends on pilot calibration and channel codec.
- Does this work on standard telephony networks?
- Yes. Sonotheia is validated for wideband, G.711, and AMR-NB telephony channels. Pilot calibration establishes channel-specific thresholds, and real-time claims are limited to evaluated pilot or demo contexts.
- What happens on a false positive?
- Explainable outputs show which sensors triggered so fraud ops can tune thresholds without black-box guesswork. Conformal abstention profiles are available for high-stakes channels.
- Is customer audio sent to Sonotheia cloud for training?
- No. Deployment defaults to on-prem or customer VPC. Audio is processed in memory and not retained.
- How does this relate to NYDFS voice deepfake guidance?
- Voice deepfakes are increasingly framed as biometric circumvention threats. Local voice liveness and texture verification provide a defensive control auditable under existing cybersecurity programs.
- What is the first step for a bank evaluation?
- Schedule a controlled pilot via /request-demo. We scope channel mix, codec profile, and acceptance metrics before any production commitment.
Ready to evaluate Sonotheia for your voice channel?
Request a demo