GGMDDC: An Audio Deepfake Detection Multilingual Dataset

Audio Deepfake Detection (ADD) is of critical concern due to its social relevance. The development of ADD solution faces the limitation of the low diversity of language resources and is restricted to a limited number of speakers. However, there is a need for a more reliable ADD system due to advancements in deepfake audio generation by the attackers. Most of the existing datasets for ADD task are limited to a language count, so their usage remains limited to restricted number of speakers. Also, in most of the existing datasets for ADD, the number of issue of some speakers’ real and deepfake exists, and others are unavailable. In this study, we implemented a configuring version of Generative Adversarial Networks (GANs), namely, High Fidelity-GAN, to create dataset, GAN-Guideded Multilingual Deepfake Detection Corpus(GGMDDC). We achieved Mean Openion Score (MOS) of 4.52 for the proposed system, which proved to be better than MOS of the existing systems.