Telephony-native voice anti-spoofing for Gnani.ai
A phone line costs published anti-spoofing models 12 to 16 EER points on codecs they never saw, and a threshold fixed on dev data does not survive the trip.
Data: REAL audio (ASVspoof 2019 LA clips, English proxy), SIMULATED phone channel (real ffmpeg codecs)
The problem
Gnani's Armour365 verifies callers by voice. A cloned voice made with TTS or voice conversion reaches it through an 8 kHz phone line, and codecs such as G.711, AMR-NB and GSM strip many of the cues a countermeasure relies on.
This build asks three questions: how much does the narrowband channel hurt a countermeasure, does a threshold fixed on dev survive it, and can an IVR flow on Gnani Prisma STT and Timbre TTS state an honest operating point.
Every clip is VAD-trimmed and loudness-normalised, then sent through 14 telephony conditions with real ffmpeg codecs. Four conditions (AMR-NB 4.75k, GSM-FR and two two-hop chains) are held out of training.
Key result
Codec augmentation holds the line: held-out-codec EER 34.42% with it vs 41.84% without (same LCNN). Best published model, SSL+AASIST: 9.15% clean, 20.97% on held-out codecs.
Equal error rate, clean vs held-out codecs (lower is better)
Comparison with baselines
Countermeasures on ASVspoof 2019 LA eval (2,560 utterances, unseen attacks A07 to A19)
| Model | EER clean | EER seen | EER held-out [95% CI] | minDCF held-out | FRR@FAR1 held-out | delta-NB |
|---|---|---|---|---|---|---|
| LFCC-GMM 8 kHz, codec-aug | 25.59 | 30.38 | 33.63 [32.29, 34.81] | 0.781 | 99.28 | +8.04 |
| LFCC-LCNN 8 kHz, codec-aug | 27.64 | 30.70 | 34.42 [33.04, 35.60] | 0.799 | 92.47 | +6.78 |
| LFCC-LCNN, no codec aug (ablation) | 26.59 | 35.17 | 41.84 [40.63, 42.86] | 0.943 | 98.45 | +15.25 |
| LFCC-GMM clean 16 kHz (wideband control) | 24.10 | 41.63 | 44.73 [43.53, 45.97] | 0.998 | 100.00 | +20.63 |
| AASIST (published, 16 kHz) | 23.65 | 37.30 | 39.22 [37.63, 40.20] | 0.855 | 92.17 | +15.57 |
| AASIST-L (published, 16 kHz) | 21.57 | 34.32 | 37.13 [35.89, 38.35] | 0.869 | 96.78 | +15.56 |
| SSL+AASIST (published, 16 kHz) | 9.15 | 18.45 | 20.97 [19.93, 21.96] | 0.572 | 71.60 | +11.83 |
| XLS-R teacher, fine-tuned at 8 kHz | 23.34 | 24.23 | 21.63 [20.03, 22.71] | 0.381 | 7.35 (FAR 27.0) | -1.71 |
| Distilled LCNN student | 26.79 | 30.58 | 34.63 [33.35, 35.86] | 0.836 | 91.17 | +7.84 |
EER, FRR in %. delta-NB = EER(held-out) minus EER(clean pool), same utterances. FRR@FAR1 uses the threshold fixed on dev at FAR 1%; for the teacher that threshold lets 27% of spoofs through on held-out codecs. SOTA and teacher rows: Modal GPU runs (about $9.42 in total, including a first teacher run that did not learn).
IVR flow on real eval clips at dev-fixed operating points (200 genuine + 200 attack clips per codec)
| Codec | Operating point | Genuine accepted | Attack accepted |
|---|---|---|---|
| G.711 (seen) | far1 | 32.5% | 11.0% |
| G.711 (seen) | min_dcf | 88.5% | 62.5% |
| G.711 (seen) | eer | 84.5% | 49.0% |
| AMR-NB 4.75k (held out) | far1 | 4.5% | 1.0% |
| AMR-NB 4.75k (held out) | min_dcf | 75.5% | 55.5% |
| AMR-NB 4.75k (held out) | eer | 57.5% | 33.5% |
CM gate only: echo STT, identity gate bypassed. No operating point meets genuine >= 90% with attack <= 5%. Median local latency about 30 ms.
Other findings
- Published SOTA does not survive the phone line intact: SSL+AASIST goes from 9.15% to 20.97% EER on held-out codecs, and AASIST (39.22%) and AASIST-L (37.13%) end up worse than the 170k-parameter codec-augmented LCNN (34.42%).
- Fine-tuning SSL+AASIST at 8 kHz with codec augmentation made it channel-flat (delta-NB -1.71, best minDCF 0.381) but not better overall: 21.63% held-out EER, inside SSL+AASIST's interval, because it lost unseen attacks such as A11 (7.16% to 49.84%).
- Distilling that teacher into the LCNN gave no gain (34.63% vs 34.42% held-out), so the shipped CPU model stays the plain codec-augmented LCNN.
- Codec augmentation is what holds the line for small models: the same LCNN without it loses 15.25 EER points on held-out codecs (41.84% vs 34.42%).
- The damage concentrates on attacks that were easy in clean narrowband: LCNN A08 2.12% to 18.86%, A09 0.22% to 9.78%, A17 4.84% to 19.00%.
- Shortcut audit: raw silence alone gives 17.54% EER; on the CM input every cue is at least 37.0%.
- Efficiency: the LCNN in ONNX has 170,401 parameters, RTF 0.0050 on one thread, 143 MB peak RAM.
What this does not show
- English proxy only (ASVspoof 2019 LA) for every ranked number. An Indic pilot (FLEURS vs IndicSynth FreeVC) fails its shortcut audit (duration alone: 16.6% EER), so it supports no model claim. No Gnani-clone result exists yet.
- Phone channels are simulated with real codecs, not recorded phone calls.
- Both CPU baselines are weak (the official 16 kHz LFCC-GMM is about 8% EER), so the codec effect is only readable on attacks they catch.
- The teacher was fine-tuned on a 5,160-utterance subset for 10 epochs and scored on full clips only (no streaming rows, not scored on the Indic pilot).
- The speaker-verification stage is a placeholder: the demo and trials bypass the identity gate.
- At a dev-fixed FAR of 1%, the CM rejects 92% of genuine callers on held-out codecs; no tested operating point is deployable as is.
- The 4 s streaming row has only 4 bona fide clips and is unreadable.
How to reproduce
From the repository root (commands from its README; Windows PowerShell unless noted).
.\run.ps1 setup # venv, CPU torch, requirements, clones AASIST and SSL_Anti-spoofing
.\run.ps1 quick # under 5 minutes: unit tests, teacher smoke, IVR demo on real clips, tables
.\run.ps1 trials # IVR flow on real ASVspoof clips at the dev-fixed operating points
.\run.ps1 report # rebuild artifacts/results.md from release/
.\run.ps1 gpu-pretrained # Modal: AASIST, AASIST-L, SSL+AASIST on every condition
Source of every number on this page: REPORT.md (state on 2026-10-24) and README.md in the build's repository.