Flagship Engine

PhantoM

Voice Deepfake Detection Engine

Telephony-grade voice deepfake detection for call centers - streaming, low latency, DD-grade evidence. Acquisition-first packaging.

Low Latency (NDA)
8kHz Telephony Grade
534K Eval Samples

At a Glance

Detect AI-generated voice in real-time over telephony connections (8 kHz, G.711/Opus). PhantoM is packaged as an IP acquisition asset with reproducible benchmarks, raw evaluation scores, and evidence-first documentation - not a SaaS product.

The Problem

Voice deepfake attacks are now a practical fraud vector in call centers and voice-authentication flows. Traditional controls struggle with compressed telephony audio, short interaction windows, and the operational cost of false positives. Buyers need reproducible benchmark evidence for narrowband (8 kHz) telephony conditions.

The Solution

PhantoM provides a streaming detection core designed for narrowband telephony audio (8 kHz) with an enterprise deployment wrapper and an evidence-first diligence package. Buyers can verify latency, benchmark methodology, and operating points under NDA using raw-score artifacts and gate reports.

Capabilities

Capabilities designed for enterprise integration.

📞

Telephony-Grade Streaming

Real-time scoring over 8kHz narrowband audio with telephony-focused preprocessing.

FAR-Anchored Decisioning

Operating points expressed in FAR/TPR terms to match call-center risk tolerances (e.g., TPR at FAR=1-2%).

Telephony Robustness Suite

Benchmark harness across codec simulation (G.711/Opus), jitter, packet loss, and noise conditions.

🔒

Deploy Anywhere

On-prem / self-hosted deployment design. No requirement to send audio outside buyer infrastructure.

Enterprise Security Wrapper

Enterprise wrapper patterns (mTLS, RBAC, audit logging) documented under NDA; implementation scope varies by integration.

Acquisition-First Packaging

Stage-gated disclosure and DD-grade artifacts designed for low-friction M&A / IP acquisition.

Evidence & Proof Points

DD-friendly artifacts and verifiable outputs for technical evaluation.

Low-latency
p95 (NDA)
Exact figures in NDA benchmark report
Benchmarked
EER (In-Domain)
Full metrics in NDA evidence pack
534K
Cross-Domain Eval
ASVspoof 2021 DF eval set
21+
Impairment Tests
Codec, jitter, noise matrix

Sample Outputs

Benchmark Report (methodology + integrity hashes) — under NDAOperating Point Table (FAR → threshold → TPR) — under NDATelephony Corruption Matrix (JSON) — under NDALatency Benchmark Report (JSON) — under NDASBOM (CycloneDX JSON) — under NDAGate Verification Reports (JSON) — under NDAProof Card (1-page summary) — teaser available on request

All samples are synthetic (no customer data, no secrets). Full evaluation evidence packs are available under NDA during technical evaluation.

Integration

Clear inputs and outputs for seamless integration into your stack.

Inputs

  • Streaming audio chunks (8 kHz narrowband or resampled)
  • Encodings: PCM_S16LE, G.711 u-law, G.711 A-law, Opus (simulation-supported)
  • Optional call metadata (session ID, timestamps)

Outputs

  • Per-chunk confidence score (human <-> synthetic)
  • Verdict: human / synthetic / uncertain
  • Latency per chunk
  • Optional audit events (JSON)

Ideal For

Best-fit buyer profiles and use cases.

Contact Center Platform Vendors

Embed voice deepfake detection as a native platform capability.

Voice Biometrics Vendors

Add synthetic voice detection to your authentication stack.

Banks & Financial Services

Protect high-value voice transactions from impersonation attacks.

Telecom Security Teams

Deploy telephony-grade fraud-risk signals at the network edge.

KYC/IDV Providers

Add synthetic-voice risk scoring to voice verification workflows.

M&A Due Diligence

Acquire DD-grade IP with reproducible benchmarks and evidence chain.

Known Limits & Scope

  • Preferred model: full IP acquisition (transfer of ownership). Exclusive licensing discussed case-by-case under NDA. No hosted SaaS option.
  • Evidence pack and sensitive artifacts shared under mutual NDA
  • No ongoing support post-acquisition; fixed-scope transition support (20-40 hours)
  • Cross-domain performance is disclosed; DF-domain retraining expected for production deployments
  • No SOC 2 / ISO certification (solo-engineered asset; architecture artifacts provided)
  • Performance degrades under severe bandpass/jitter/noise conditions (documented)
  • Not a full identity verification or liveness product; provides synthetic-voice detection signal for integration into existing workflows
  • Benchmarked on ASVspoof 2021 DF dataset (Open Data Commons Open Database License v1.0); dataset not redistributed; third-party license inventory available under NDA

Ready for a Deep Dive?

Schedule a 30-minute technical walkthrough to see PhantoM in action and discuss integration options.