Audio deepfake detectors are usually evaluated on clean digital audio, but real calls pass through acoustic capture, device processing, and transmission coding that alter the signal. To measure this gap, the authors introduce RealComm, a dataset that pairs digital speech with recordings of the same utterances transmitted through actual mobile calls. The acquisition design includes acoustic and wired injection, handset and speakerphone operation, and caller-side denoising, with both bona fide and synthetic speech recorded under matched conditions.
Benchmarking existing detectors on this data shows substantial degradation after call transmission, even for models that perform well on digital audio. Acoustic injection is particularly difficult. The authors also report that measured quality degradation is associated with poorer detection across recording conditions, while high-frequency attenuation alone does not explain the difficulty.
The paper proposes RealComm-Aug, a staged waveform augmentation pipeline that simulates acquisition and transmission effects, and investigates it alongside real-call training. These two approaches provide complementary benefits: combining them yields the lowest call-domain error for all three evaluated detectors. For Teffic-Audio, the equal error rate falls from 37.94% to 22.37%. The dataset is split into RealCommBench for evaluation and RealCommTrain for adaptation, with access via the project repository.