An Improved System for Partially Fake Audio Detection Using Pre-trained Model
摘要
The technology of speech synthesis and conversion has made good progress with the development of deep learning. However, such technology can also do harm to information security and may be applied for illegal uses. Therefore, researchers have conducted a lot of research on the task of speech deep forgery detection recently. Corresponding to this, the Audio Deep Synthesis Detection Challenge 2023 (ADD 2023) is held. In this paper, we propose a fake audio detecting system using a pre-trained model, focusing on partially fake audio detection tasks. We have presented our models to the ADD 2023 challenge. In the final competition, our system got a score of 0.4855 in the manipulation region location track.