Many voice deepfake detectors perform exceptionally well in laboratory testing on high-fidelity, studio-recorded audio. However, when these same systems are deployed on real-world telephony channels, their accuracy can degrade significantly.
The reason is simple: telephony networks compress audio aggressively. Without codec-specific validation and calibration, deepfake detection systems fail silently or generate high false-alarm rates.
The Auditory Destruction of Lossy Networks
Telephony systems rely on codecs to minimize bandwidth. In standard telecommunication, the dominant protocols include:
- G.711 (u-law/a-law): Standard landline and legacy VoIP compression, limiting frequency bandwidth and quantizing sample amplitudes.
- AMR-NB (Adaptive Multi-Rate Narrowband): The standard codec for mobile cellular networks, which operates at very low bitrates and heavily compresses speech signals.
These codecs strip high-frequency content and introduce compression artifacts.
Because many black-box AI deepfake detectors are trained on studio-grade, high-fidelity datasets, they learn to rely on high-frequency spectral cues to identify synthetic voices. When that audio is transmitted over a cell network using AMR-NB, the codec completely erases those high frequencies. The detector is left blind, unable to distinguish a sophisticated clone from a genuine caller.
Calibrating for Codec Survival
To build resilient voice fraud controls, detection systems must be validated on compressed signals. This means:
- Physics-Based Feature Extraction: Focusing on low-frequency signal dynamics, such as spectral envelope trajectories and excitation phase patterns, which survive lossy compression.
- In-Situ Calibration: Tuning detection thresholds specifically for the codec of the target channel (e.g., separate baselines for G.711 vs. wideband networks).
Benchmarking on simulated codec chains can identify limitations under those transforms. It does not establish performance on live telephone traffic, which adds handset, network and transcoding effects that need measurement in place.
Detection accuracy claims only mean something once real telephony channel conditions are accounted for. See how Sonotheia's voice and channel measurements are benchmarked on simulated G.711 and AMR-NB chains. Live telephone traffic has not been validated by those simulations.