ConvNext Based Neural Network for Audio Anti-Spoofing

Automatic speaker verification (ASV) has been widely used in the real life for identity authentication. However, with the rapid development of speech conversion, speech synthesis algorithms, ASV systems are vulnerable for spoof attacks. In recent years, there have many works about synthetic speech detection, researchers had proposed a number of anti-spoofing methods based on hand-crafted features to improve the detection accuracy and robustness of ASV systems. However, using hand-crafted features rather than raw waveform would lose certain information for anti-spoofing, which will reduce the detection performance of the system. Inspired by the promising performance of ConvNeXt in image classification tasks, we revise the ConvNeXt network architecture accordingly for spoof attacks detection task and propose a light weight end-to-end anti-spoofing model. By integrating the revised architecture with the channel attention block and using the focal loss function, the proposed model can focus on the most informative sub-bands of speech representations to improve the anti-spoofing performance and the difficult samples that are hard for models to classify. Experiments show that our proposed best single system could achieve an equal error rate of 0.75% and min-tDCF of 0.0212 for the ASVSpoof2019 LA evaluation dataset, which outperform the state-of-the-art systems.