Adaptive data-driven audio network for speech deepfake detection

Abstract

Automatic speaker verification (ASV) systems are often vulnerable to spoofed attacks, particularly unseen attacks. Among the audio deepfake detection algorithms, AASIST and RawNet2 are primarily used. These methods employ the Sinc filters as encoders to learn an efficient filter bank. The key limitations include the model's lack of generalization to different datasets, and the computational cost is high compared to simpler feature models. In this work, we propose the adaptive data-driven audio network. This network identifies the sparse coefficients for the input audio patches and updates the filters to better align with the data. It allows the model to extract more relevant and complex features of audio signals. The network reduces computational time to one-third of the network algorithms with Sinc filters. Experimental results on English, Chinese, and Cantonese audio deepfake detection demonstrate that the network outperforms other typical algorithms, such as AASIST and RawNet2. The network offers greater stability and robustness.