Speech Spoofing Detection Based on One-Class Residual Attention Network

Detection of spoofed speech to ensure the authenticity and integrity of speech is of great significance to voice authentication systems, judicial evidence collection, and social stability. Recently, one-class learning has shown promising performance in speech spoofing detection. This paper proposes a one-class residual attention network to address overfitting caused by feature distribution mismatch and increased depth in residual networks. Specifically, a one-class classification approach is used in a multi-scale residual neural network (Res2Net). The attention mechanism is applied to Res2Net, and different weights are given to the speech features of different time frames, so that the attention can be effectively paid to the spoofed speech. A one-class loss function is used for training the model, which effectively improves the detection accuracy. The proposed one-class residual attention network is evaluated on the ASV spoof 2019 dataset, and an EER of 1.72% and a min t-DCF of 0.045 are obtained, which outperform the baseline methods.