Call
—
0%
—
Response actions
Per-model scores
Reason codes
Voice Security · Node BLR-07
Run an analysis — each model's solo verdict vs. fusion appears here.
Verdicts appear here, newest first.
Do our probabilities mean what they say? Reliability curve on the held-out test set, and the low-confidence band we route to a human instead of auto-deciding.
Each lane is a real concurrent call hitting the detection engine. Click any completed call for its full report.
Start the load to see calls stream in.
Global anti-spoofing corpora, real commercial clones, and real Indian speech in 10 languages.
| Approach | Clean EER | Real-world | Speed | Best for |
|---|---|---|---|---|
| Single acoustic (LFCC) | 6.2% | 19% | Fast | Baseline |
| Single wav2vec2 | 2.9% | 14% | Medium | Generalisation |
| Fusion head (fast) | 2.1% | 13% | <1s | Real-time |
| VoxShield fusion | 1.62% | 11.6% | ~3s | High-assurance |
| Solution | Voice? | Explainable | On-prem | India-tuned | Focus |
|---|---|---|---|---|---|
| VoxShield | Yes | Reason codes | Yes | 10 langs | Bank call fraud |
| Pindrop | Yes | Partial | Cloud | US | Call-centre auth |
| Reality Defender | Yes | Partial | Cloud | No | Media deepfakes |
| Nuance Gatekeeper | Biometric | No | Hybrid | No | Voice biometrics |
| Caller-ID / CNAP | Number | Yes | Telco | Partial | Number reputation |