WaveFake: A data set to facilitate audio DeepFake detection

The main purpose of this data set is to facilitate research into audio DeepFakes.<br> These generated media files have been increasingly used to commit impersonation attempts or online harassment. The data set consists of 88,600 generated audio clips (16-bit PCM wav).<br> All of these samples were generated by four different neural network architectures: MelGAN Parallel WaveGAN Multi-Band MelGAN WaveGlow Additionally, we examined a bigger version of MelGAN and investigated a variant of Multi-Band MelGAN that computes its auxiliary loss over the full audio instead of its subbands. <strong>Collection Process</strong> For WaveGlow, we utilize the official implementation (commit 8afb643) in conjunction with the official pre-trained network on PyTorch Hub.<br> We use a popular implementation available on GitHub (commit 12c677e) for the remaining networks.<br> The repository also offers pre-trained models.<br> We used the pre-trained networks to generate samples that are similar to their respective training distributions, LJ Speech and JSUT.<br> When sampling the data set, we first extract Mel spectrograms from the original audio files, using the pre-processing scripts of the corresponding repositories.<br> We then feed these Mel spectrograms to the respective models to obtain the data set. This data set is licensed with a CC-BY-SA 4.0 license. This work was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy -- EXC-2092 CaSa -- 390781972.