ScholarWorks@중앙대학교

초록

Recently, significant research efforts have been directed toward enhancing the performance of voice spoofing detection systems by employing neural vocoders as a data augmentation technique. In this study, we investigated the impact of training spoofing detection models with spoofed speech data generated through neural vocoders. To this end, we constructed training datasets by applying four different pre-trained vocoders to the data from the Automatic Speaker Verification Spoofing and Countermeasures (ASVspoof) Logical Access (LA) track. We employed a Self-Supervised Learning (SSL)-based Countermeasure (CM) model and evaluated its detection performance across four distinct datasets. For the ASVspoof 2019 LA dataset—representing the training domain—performance gains were observed in certain vocoder configurations; however, these improvements were inconsistent across various vocoder types or their combinations. Conversely, substantial performance improvements of approximately 27\% and 47\% were achieved on external evaluation datasets, specifically WaveFake and In-the-Wild, which were not utilized during training. These findings emphasize the importance of appropriate vocoder selection in enhancing the generalization capability of spoofing detection models and demonstrate that vocoder-generated spoofed speech can significantly boost model performance.

키워드