Active Acoustic Probes for AI Voice Call Detection
TL;DR
An active acoustic probe is proposed that verifies whether the returned audio traversed a genuine physical speaker-to-microphone path, and yields a simple, explainable signal-processing rule calibrated only on normal physical-path recordings.
Abstract
AI voice calls are making it increasingly practical to impersonate people during live audio conversations, whether through voice cloning, real-time conversion, or autonomous voice agents. Current defenses remain mostly passive: they inspect received audio for artifacts created by current generators, which locks defenders into an arms race structurally favoring the attacker. Active defenses invert this dynamic by injecting fresh evidence during the call, but they typically require the remote speaker to cooperate by performing a challenge task. We propose an active acoustic probe that requires no such cooperation: it verifies whether the returned audio traversed a genuine physical speaker-to-microphone path. During a call, the verifier sends a brief high-frequency acoustic challenge and checks whether it survives in the returned audio. Normal physical paths preserve the challenge, whereas the tested voice-conversion paths suppress it. This yields a simple, explainable signal-processing rule calibrated only on normal physical-path recordings. The probe is correctly detected in 99% of trials across diverse consumer audio form factors, from open laptop/phone playback to ear-worn headphones and earbuds, while all 200 outputs from four tested speech-to-speech and voice-conversion configurations fail the detector.
