Bengali Real and Deepfake Audio Dataset for Deepfake Detection

Synthetic speech or audio deepfakes are increasingly threatening to information veracity and\u2002public trust. Though deepfake detection has seen a lot of interest, the scarcity of open-source datasets for Bengali speech hampers progress\u2002in this field. To fill this gap\u2002we present the Bengali Real and Deepfake Audio Dataset, which is a curated repository of real and fictitious speech data in Bengali language that can facilitate deepfake detection research and speaker forensics. The dataset consists of\u20024000 authentic and 4000 deepfake speech clips from 10 subjects in the height standardized at a sampling rate of 24kHz. The deepfake samples were generated using FreeVC Voice Conversion model, with realistic acoustic properties. The goal\u2002of this dataset is to offer a much-needed resource for the research community, in order to engage and tackle deepfake audio detection system in a low-resource language such as Bengali.