Abstract
Advancements in deepfake audio generation are increasingly endangering the reliability of speech biometrics, remote verification processes, and various digital messaging infrastructures. In this regard, the development of effective models that can detect synthetic speech with high accuracy is one of the current scientific tasks. The purpose of this study is to develop an MSCA-Net (Multi-Scale Convolutional Attention Network) architecture for detecting audio deepfake based on Mel-spectrograms and evaluate its effectiveness. The proposed model provides effective extraction of spectral features characteristic of synthetic speech by combining multi-scale convolutional blocks with the CBAM (Convolutional Block Attention Module) attention mechanism. The architecture was trained and evaluated on the ASVspoof 2019 Logical Access (LA) and Fake-or-Real 2 seconds (FoR-2sec) datasets. According to the experimental results, the model demonstrated 94.10% Accuracy, 95.72% Precision, 94.10% Recall, 94.56% F1-score and 4.90% Equal Error Rate (EER) on the ASVspoof 2019 LA dataset. On the FoR-2sec dataset, 92.30% Accuracy, 92.61% Precision, 92.30% Recall and 92.45% F1-score were achieved, respectively. The conducted comparative analysis showed that the proposed architecture is competitive with state-of-the-art methods. The obtained results prove that the MSCA-Net model can reliably detect synthetic speech and can be used in speech biometrics, digital content verification and cybersecurity systems.