Shipped: VoiceGuard: fake-voice detection on phone lines

Can a cheap speech deepfake detector survive a phone line? VoiceGuard trains small classifiers on frozen WavLM features from ASVspoof 2019 and tests them on a simulated phone line, phone noise, Opus voice compression and 31,779 real-world In-the-Wild clips.

What I found:

• The middle WavLM layer (6) is the most robust: 1.7% EER on clean audio, 2.2% on a phone line, 18.0% on real-world clips.
• Calibration fitted on clean audio breaks when the channel changes.
• On real-world audio, most errors are genuine voices flagged as fake, so the demo shows an "uncertain" band.
• Zero-padding short clips in a batch silently changed scores; fixed with equal-length batching.

The browser demo runs entirely on your device; audio never leaves it. Code, results and tests are open (MIT). Preprint: doi.org/10.5281/zenodo.24141726

Research prototype, not a forensic tool.