A Synthetic Speech Detection Model Combining Local-Global Dependency

Synthetic speech detection is an essential task in the field of voice security, aimed at identifying deceptive voice attacks generated by text-to-speech (TTS) systems or voice conversion (VC) systems. In this paper, we propose a synthetic speech detection model called TFTransformer, which integrates both local and global features to enhance detection capabilities by effectively modeling local and global dependencies. Structurally, the model is divided into two main components: a front-end and a back-end. The front-end of the model uses a combination of SincLayer and two-dimensional (2D) convolution to extract high-level feature maps (HFM) containing local dependency of the input speech signals. The back-end uses time-frequency Transformer module to process these feature maps and further capture global dependency. Furthermore, we propose TFTransformer-SE, which incorporates a channel attention mechanism within the 2D convolutional blocks. This enhancement aims to more effectively capture local dependencies, thereby improving the model’s performance. The experiments were conducted on the ASVspoof 2021 LA dataset, and the results showed that the model achieved an equal error rate (EER) of 3.37% without data augmentation. Additionally, we evaluated the model using the ASVspoof 2019 LA dataset, achieving an EER of 0.84%, also without data augmentation. This demonstrates that combining local and global dependencies in the time-frequency domain can significantly improve detection accuracy.