A hybrid Cochleagram–HuBERT capsule network framework for generalizable deepfake speech detection with a Turkish call-center dataset

Rapid advances in text-to-speech, voice conversion, and neural vocoder technologies have made synthetic speech increasingly difficult to distinguish from genuine human speech, creating a critical security threat for automatic speaker verification systems, call-center-based identity verification processes, and digital forensics applications. Although many methods have been proposed for audio deepfake detection, a substantial portion of existing approaches remains dependent on English-dominant benchmark datasets and limited attack types, resulting in insufficient generalization under hybrid scenarios involving different languages, data sources, and generation methods. To address this limitation, this study proposes a dual-branch Capsule Network architecture that combines Cochleagram-based biological auditory representations with HuBERT-based self-supervised speech representations. In the first branch, speech signals are transformed into Cochleagram representations and classified using a two-dimensional CapsNet. In the second branch, 1024-dimensional representations extracted from the eighth Transformer layer of HuBERT-Large are processed using a one-dimensional CapsNet. The spoofing scores produced by the two branches are then combined through mean score-level fusion to generate the final decision. Another major contribution of this study is the construction of a new deepfake audio dataset for Turkish call-center scenarios. The dataset consists of 2758 mono WAV files at 16 kHz and 16-bit PCM, including 2008 bona fide speech recordings collected from 87 speakers and 750 spoofed speech recordings generated using SpeechT5, MMS, and Chirp3. Experimental results show that the proposed hybrid model achieves 97.64 ± 2.87% accuracy and 0.155 ± 0.141% EER on the Turkish dataset, and 98.68 ± 1.55% accuracy, 97.94 ± 2.47% precision, and 0.236 ± 0.119% EER in the hybrid scenario where Turkish data are evaluated together with international TTS samples. In addition, the model obtains 0.15% EER and 0.004 min-tDCF on ASVspoof2019 LA, 0.37% EER and 0.01 min-tDCF on ASVspoof2019 PA, and 6.31% EER in the Fake-or-Real/Speech DF Arena comparison, producing lower error rates than many recent methods in the literature. These findings demonstrate that the complementary use of Cochleagram and HuBERT representations provides a more stable and generalizable audio deepfake detection system against variations in language, attack type, and data source.