Reality Defender

Nothing here exists without models. Distinguishing a cloned voice from a real one during a live call is a pure inference problem with no rules based fallback, and the company describes an ensemble of models rather than a single detector, which is the standard defence against any one model failing against a new generator. Apply the removal test and there is no residual product at all, not even a degraded one.

The product flags rather than acts, which is the right posture, and the institution decides what a flag triggers within its own fraud process. Beyond that the design is undescribed. Nothing public sets out whether output is a binary verdict or a score, what thresholds are configurable, whether a flagged call is blocked, escalated or merely annotated, or how an agent is meant to act on a synthetic voice alert mid conversation with someone who may be a real customer.

No accuracy figure, detection rate, false positive rate or evaluation methodology is published anywhere located in this pass, leaving high accuracy as the entire disclosure. This field is unusual in having independent public benchmarks for synthetic speech detection that a vendor can enter and be scored on, in the same way biometric vendors submit to government face evaluations, and no evidence of participation was found. For a product whose single function is a detection judgement, publishing nothing about how often that judgement is right is the central gap.

A pattern worth naming runs through the published material: the threat is quantified exhaustively and the product barely at all. Losses per breach, incident prevalence, projected sector losses and phishing growth are all cited with named third party sources, while the product's own performance is described only as maintaining high accuracy.

The strongest evidence is a global tier one bank deployment protecting high net worth clients through the contact centre, said to process thousands of calls daily, published as a downloadable case study with the institution unnamed, alongside analyst recognition naming the company the one to beat in its category.

Ensemble design is a genuine safety choice, since it avoids a single detector becoming the single point of failure when a new generator appears. That appearance is the whole problem in this category: detection is an arms race in which a model's accuracy decays as generation improves, so retraining cadence against new voice cloning tools is the central stewardship question for this vendor specifically. Nothing public describes that cadence, the adversarial testing regime, whether analysed customer audio is used for retraining, or which providers are involved.

Real time analysis of live calls means customer voice audio flows through the detection layer, and voice carries a specific legal weight that behavioural signals do not: voiceprints are explicitly enumerated in the strictest state biometric privacy statutes, one of which carries a private right of action, so processing customer speech sits inside a regime that behavioural biometrics arguably escapes. Nothing public addresses that, and no privacy framework, retention position for analysed audio, or subprocessor disclosure was located.

No trust centre, enumerated certification list, attestation scope or audit period was located in this pass. A tier one bank running the platform across its contact centre would have required attestations before that deployment, so the published record understates the control environment, and the grade records what an outside buyer can verify rather than a judgement on the controls themselves.

Reality Defender supplies technology and holds no licence, which is expected, but its regulatory grounding is thinner than the compliance vendors in this index because it is a security product rather than an obligation driven one. Published material engages with the policy environment, citing a Federal Reserve governor on generative fraud risk, without mapping the product to any named requirement. No supervisory instrument is identified, no formal admission process has been passed, and nothing states how detection output would be evidenced to an examiner.

A voice deepfake detector decides whether speech sounds genuine, and everything that makes a real voice deviate from the training distribution pushes toward a false positive: a strong regional or foreign accent, a non native speaker, a speech impairment, an older or ill voice, a poor line. The consequence is that a real customer is treated as an impersonator and denied access to their own account, and it falls hardest on the people least equipped to argue their way past it. No per demographic false positive rates, no accent or language coverage disclosure, and no accessibility analysis were located.

Both error directions carry real cost and neither is addressed. A missed detection lets an impersonated transfer proceed, and the institution absorbs it. A false positive tells a bank that a genuine customer is synthetic, and that customer is locked out of their own money with no way to know an AI judged their voice, let alone challenge it. No accuracy guarantee, no remediation commitment, no published error rate and no correction route were located for either party.

The ensemble is presented as proprietary and comprehensive, which implies models built and owned in house rather than resold, and that is a genuinely short chain if accurate. It is implied rather than stated. No model providers are named, no training data sources are described, no subprocessor list is published, and nothing identifies where live call audio is processed, which is the disclosure a bank would need before routing customer speech through a third party in real time.

The integration posture is additive and aimed at the channels where the attacks land, running inside contact centre infrastructure to score live calls, inside web conferencing to check participants, and alongside existing liveness and verification checks rather than replacing them. That matters commercially, because an institution can adopt detection without re procuring telephony or identity infrastructure. What was not located is the detail: no named contact centre, conferencing or verification platform partners, no public developer documentation, and no marketplace listings.

Delivery is cloud hosted with real time processing, which means live customer call audio leaves the institution's environment as it is spoken. That makes residency an immediate operational question rather than a storage one, and it is unanswered: no hosting regions, in country processing options, transfer mechanisms or subprocessor chain were located, and nothing states whether audio is retained after a verdict is returned or discarded in flight.

No rates, tiers, billing unit or minimum were located in this pass. The unit question is the interesting one and is unanswered: real time detection across a contact centre could plausibly be charged per call, per minute of audio analysed, per seat or per enterprise, and each produces a very different cost curve for an institution whose call volume is not discretionary.

Financial services is addressed as one solution area among several rather than as the design centre, with the company also serving government, media and general enterprise, and the finance material concentrates on large institutions with contact centres and high net worth client bases.

Within finance the coverage is thin by institution type: no dedicated treatment of credit unions, insurers, capital markets or payments firms, and the use cases cluster around three channels, contact centre voice, video conferencing and identity verification.