홈 › Papers

Model Attribution and Detection of Synthetic Speech via Vocoder Fingerprints

As speech generation technology advances, so do the potential threats of misusing synthetic speech signals. This work tackles three tasks: (1) single-model attribution in an open-world setting corresponding to the task of identifying whether synthetic speech signals originate from a specific vocoder (which requires only target vocoder data), (2) model attribution in a closed-world setting that corresponds to selecting the specific model that generated a sample from a given set of models, and (3) distinguishing synthetic from real speech. We show that standardized average residuals between audio signals and their low-pass or EnCodec filtered versions serve as powerful vocoder fingerprints that can be leveraged for all tasks achieving an average AUROC of over 99% on LJSpeech and JSUT in most settings. The accompanying robustness study shows that it is also resilient to noise levels up to a certain degree.

Code (0)

등록된 구현이 없습니다.

Methods 이 논문이 사용한 방법론

Similar Papers 제목 키워드 기반