Partially Fake Audio Detection Based on Mamba and Tensor Feature Fusion

In recent years, models related to Artificial Intelligence Generated Content (AIGC) have advanced rapidly, opening up new possibilities for generating realistic speech. However, if misused, this technology can pose significant risks to information security. Consequently, the task of deepfake audio detection has emerged. Within this domain, the Manipulation Region Location (MRL) task specifically aims to identify the manipulated segments of speech, offering greater precision and facilitating downstream tasks such as intent analysis. In this paper, we first investigate the performance of various feature types in the MRL task to determine which features exhibit the best generalization ability. Then we propose a tensor-based feature fusion strategy to effectively capture the interrelationships among different features and produce a more representative fused feature. Furthermore, leveraging the strong temporal modeling capabilities of Mamba, we incorporate it into our framework. To the best of our knowledge, this is the first work to introduce Mamba into the MRL task. The model is trained on the ADD2023 dataset. Experimental results on the test set demonstrate that our approach achieves a 37.5% improvement in performance over the baseline system and outperforms other compared models.