The article discusses the challenge of detecting replay attacks, which are simple but effective voice spoofing attacks. A method is proposed that uses audio compression to assist feature extraction, which is crucial for detecting these attacks. The technique preserves content and speaker information for transmission, while the information lost after decompression is expected to contain channel noise added during the replay process, which is key for detection. The effectiveness of the proposed method is confirmed through comprehensive experimentation with data augmentation techniques and three classifiers.
Publication date: 10 Oct 2023
Project Page: Unavailable
Paper: https://arxiv.org/pdf/2310.05813
