The voice deepfake detector that holds up on real phone lines.
DeepBlocker Detect is the same telephony-tuned model that powers Real-Time Protection, available as a hosted API for fraud, KYC, and identity-verification platforms. Reproducible evaluation, published model card, ~1 s warm inference per call.
Who it's for
- Fraud-prevention platforms scoring inbound calls.
- KYC and identity-verification vendors validating live voice.
- Bank in-house anti-fraud teams adding a voice layer to existing rule engines.
- Telecoms operators offering deepfake-defence as a network service.
Inputs and outputs
| Input audio formats | G.711 µ-law / G.711 A-law / G.722 (8 kHz, mono or stereo) |
| Window length | 4 s |
| Output — fake_prob | float in [0, 1] |
| Output — band | REAL (< 0.71) · UNCERTAIN · FAKE (> 0.96) on phone lines |
| Output — per-channel breakdown | For stereo (caller leg + assistant leg) |
Detection performance
| AI callers caught on real phone calls (previous model 86.2%) | 93.6% |
| AI callers that slip through as real (previous model 8.5%) | 3.2% |
| Real callers wrongly flagged (unchanged) | 3.5% |
| ROC AUC on the same real calls (previous model 0.963) | 0.986 |
| AI callers still caught under office noise at 10 dB (previous model 73.2%) | 80.8% |
| ElevenLabs clones over G.711, mean score | 0.953 |
Real-call figures: 207 phone calls taken off the carrier (113 real callers, 94 AI callers), both models scored on the same calls the way production scores them. Noise and ElevenLabs figures: held-out Reporting set, SHA-256 pinned. Known limitations in the model card at /model-card.
Latency, throughput, deployment
- ~1 s warm inference per stereo call. ~17 s cold start.
- Hosted on serverless GPU. Scale-to-zero.
- Calibrated thresholds versioned via secrets — no silent threshold drift.
- Continuous evaluation hooks against production drift.
What the model is honest about
- OpenAI Realtime / GPT-4o-voice currently lands in UNCERTAIN, not FAKE. Documented; on the v6 roadmap.
- Miss rate 10.2% — most slip into UNCERTAIN, not REAL. Conservative by design.
- English only. FR / ES / DE on the v6 roadmap.
- TTS providers we did not train on (Bark, Coqui XTTS-v2, Vall-E, custom voice clones) are not measured.
Reproducibility
The base model (XLSR-Mamba-LA) is MIT-licensed and published with paper. Our fine-tune procedure, augmentation chain, training hyperparameters, and Reporting set SHA-256 are documented in the model card. Available under NDA for technical evaluation.
Commercial
Tiered per-call pricing with sandbox access for technical evaluation. Volume committed pricing for fraud and KYC platforms. Contact sales for a tailored quote and a sandbox key.