One-class network leveraging spectro-temporal features for generalized synthetic speech detection

Synthetic speech attacks pose significant threats to Automatic Speaker Verification (ASV) systems. To counter these, various detection systems have been developed. However, these models often struggle with reduced accuracy when encountering novel spoofing attacks during testing. To address this issue, this paper proposes a One-Class Network architecture that leverages features extracted from the log power spectrum of the F0 subband. We have developed an advanced spectro-temporal enhancement module, comprising the Temporal Correlation Integrate Module (TCIM) and the Frequency-Adaptive Dependency Module (FADM), to accurately capture F0 subband details. TCIM captures crucial temporal dynamics and models the long-term dependencies characteristic of the F0 signals within the F0 subband. Meanwhile, FADM employs a frequency-adaptive mechanism to identify critical frequency bands, allowing the detection system to conduct a thorough and detailed analysis. Additionally, we introduce a KLOC-Softmax loss function that incorporates the KoLeo regularizer. This function promotes a uniform distribution of features within batches, effectively addressing intra-class imbalance and aiding balanced optimization. Experimental results on the ASVspoof 2019 LA dataset show that our approach achieves an equal error rate (EER) of 0.38% and a minimum tandem detection cost function (min t-DCF) of 0.0127. Our method outperforms most state-of-the-art speech anti-spoofing techniques and demonstrates robust generalizability to previously unseen types of synthetic speech attacks.