Supervisory documentation
Forensic reports that record what was measured, which thresholds applied, and the rationale, written so a supervisor, auditor, or SAR preparer can follow them.
sonotheia
Built for compliance, fraud, and risk teams.
Voice decisions, defensible under scrutiny.
Glass box, not black box. Physics-based evidence for review.
Open demoServices
Supervisory documentation. Audit trails. Evidence for review.
Records a reviewer can inspect: what was measured, which thresholds applied, and why the outcome followed.
Voice fraud governance
Ingest audio and get a forensic risk event with acoustic tags and explainable reason codes, not just a score.
Forensic analysis dashboard
Multi-codec spectral and temporal analysis so experts can validate detections and produce review-ready reports.
Technology
The measurements behind the analysis
Detection is only the first layer. Each output is paired with the measurements behind it and with governance artifacts written for review.
Detection layer
Three measurements, each reported with its limits
Each measurement tracks a different physical property of speech. They run in parallel and converge into one documented decision path.
Breath and phrasing
Speech duration
Measures how long continuous voiced speech runs between pauses and compares it with the range people produce when breathing normally. An unusually long run is reported as an observation with its threshold for a reviewer to weigh. It is not proof that the speaker is synthetic.
Room and line
Background consistency
Tracks whether the room sound and background noise stay consistent across the call. A shift can come from splicing or generation, and it can also come from a handset change or a codec switch, so the measurement is reported alongside those alternatives rather than as a conclusion.
Pitch and loudness movement
Voice changes over time
Measures how pitch and loudness move over time and whether that movement keeps the variability of natural speech. Over-regular movement is reported as an observation with its threshold. It says nothing about who is speaking.
These are bounded observations about the audio signal. They do not establish human origin, identity, fraud, intent, or authorization. Each is reported with its threshold and known limits so a reviewer can weigh it.
Governance layer
Supervision built in, not bolted on
The governance layer turns measurement output into traceable decisions, from first flag through review documentation.
Explainability
Decision trace
An analysis result links to the measurements and documented thresholds behind it so investigators can follow the logic step by step.
Reproducibility
Calibration & baselines
Frozen baselines and versioned calibration record what was in production when a decision was made.
Review support
Counter-hypothesis evaluation
Listed alternative explanations, including codec degradation, bandwidth limits, and environmental noise, are recorded with a flagged result before escalation.
Design principles
Intentional constraints
Product constraints are governance features. What we refuse to do is as important as what we automate.
No neural networks in the decision path
Classification uses inspectable linear models on hand-crafted features, with thresholds that can be documented and reviewed.
Documented measurement boundaries
Measurements describe properties of audio. Confirm data handling for each deployment.
No confidence scores without evidence
Outputs focus on measured anomalies and supporting rationale, not standalone confidence percentages.
No training on the attacks we detect
Physics-based features target properties of natural speech rather than signatures of one specific synthesizer family.
Validation
Calibrated against public benchmarks
Performance is validated on ASVspoof benchmarks across wideband, G.711 μ-law, and AMR-NB conditions. New features must pass a Do-No-Harm gate before release.
Governance artifacts are generated from the same evidence path used for detection decisions.
Trust & Governance
Clear boundaries on data and decisions
We avoid legal overclaims by design. The platform is built to reduce exposure, improve audit readiness, and produce records your teams can review and defend.
What we log and why
We log run metadata, thresholds, sensor outputs, and decision rationale so an analysis result can be traced to the measurements behind it. That record supports supervision, internal escalation, and external review.
Confirm data handling
Audio storage, analysis-result retention, access controls, and subprocessors depend on the deployment. Confirm these terms for your environment before use.
Governance primitives
- Baseline locking. Freeze production baselines so historical decisions are reproducible.
- Drift tracking. Versioned calibration history records what changed, when, and why.
- Decision trace. An analysis result links to the measurements and thresholds behind it.
- Counter-hypothesis review. Listed alternative explanations (codec, environment, noise) are recorded with a flagged result before escalation.
Context
Why voice fraud becomes a governance problem
Most tools can flag suspicious audio. Fewer can explain a decision in a way an examiner, auditor, or legal team can reliably review.
In high-stakes channels, one false positive can freeze a transfer and trigger immediate scrutiny: what was measured, which thresholds were applied, and why that conclusion was reasonable.
Built for regulated voice channels
We focus on organizations where voice decisions authorize money movement or account control: community and regional banks, credit unions, broker-dealers, RIAs, and family offices.
Regulatory references
- FINRAThe 2026 FINRA Annual Regulatory Oversight Report (December 2025) lists voice clones among generative AI fraud methods and asks firms to assess how third-party vendors use generative AI. It does not prescribe a detection technology. FINRA 2026 Annual Regulatory Oversight Report
- FinCENAlert FIN-2024-Alert004 (November 13, 2024) requests that institutions reference the alert in SAR field 2 and in the narrative using the key term FIN-2024-DEEPFAKEFRAUD. FinCEN Alert FIN-2024-Alert004
- EU AI ActArticle 50 of Regulation (EU) 2024/1689 sets transparency duties for providers and deployers of certain AI systems, including machine-readable marking of synthetic audio and disclosure of deep fakes. It applies from 2 August 2026 and is not a mandate for detection tools. Regulation (EU) 2024/1689 on EUR-Lex
- ColoradoSB26-189, signed May 14, 2026, repeals and reenacts SB24-205 with requirements for automated decision-making technology used in consequential decisions, a category that includes financial or lending services. Developer documentation duties start January 1, 2027. Whether it applies to a given institution is a question for counsel. Colorado SB26-189
These are references to primary sources, not legal advice and not a statement that any Sonotheia output satisfies a regulatory requirement.
The operating goal: defensible decisions under scrutiny
Sonotheia pairs detection with decision governance so an analysis result can be traced to the measurements behind it, then reviewed and challenged with evidence rather than confidence scores alone.
Founding Team
Founded by a compliance executive and an audio engineer. We build for audit committees and examiners — not just detection accuracy.
Doron Reizes
CO-FOUNDER & PRESIDENT
Doron Reizes is an audio engineer and educator with professional audio work since 2005, including audio post-production at Sony Music Studios. He has published sound credits in television, film, and concert projects. He is a Professor and Course Director at Full Sail University and leads Sonotheia's detection-science research.
Leads detection R&D and the firm's physics-based approach to voice integrity.
View full profile →Alexander Forostenko
CO-FOUNDER & CEO
Financial services and regulatory leader with 15+ years navigating SEC, FINRA, CFTC, and Federal Reserve expectations. Licensed attorney; former Managing Director at Charles Schwab. Senior roles at SVB, Citizens, and Morgan Stanley.
Leads operations and client relationships, ensuring programs remain defensible under scrutiny.
View full profile →FAQ
Frequently asked questions
Practical questions from compliance, risk, and fraud teams evaluating deployment.
- How is this different from detection vendors?
- Most vendors stop at scores and APIs. Sonotheia adds a governance layer: decision trace, calibration history, and review-ready artifacts.
- How is audio and analysis data handled?
- Data handling depends on the deployment. Confirm audio storage, result retention, access controls, and subprocessors for the environment you use. How to document synthetic voice evidence
- How do evidence packages support SAR and regulatory filings?
- Evidence packages give your compliance team structured measurements, thresholds, and rationale to reference when preparing SAR narratives or regulatory responses. Your team decides what to file.
- What does explainability mean in practice?
- An analysis result links to the measurements and documented thresholds behind it. You can see which sensors contributed, what rules were triggered, and the rationale recorded for the outcome.
- How do you handle drift and updates?
- Baseline locking and calibration history preserve reproducibility. Updates are versioned so teams can show exactly what was active at any point in time.
- Are you replacing existing vendors?
- Usually no. Sonotheia is designed to complement existing detection tools with governance artifacts and supervisory documentation.
Next step
Launch with a controlled pilot.
We will scope your voice channels, define governance artifacts, and map an evidence workflow your risk and compliance teams can review before rollout.
Back to top