Convolutional Speech Verification: Detecting Synthetic Speech Using Deep Learning

Today, with the unlawful activities and manipulations carried out by deep fakes, fake speech detection has become highly significant. Sound synthesis produced by specific deep learning algorithms is known as a “deep fake.” While approaches based on deep learning have demonstrated improved performance and encouraging outcomes, they have significant dependence on training data and require careful selection of hyperparameters, which poses a barrier to generalization. Literature proves that Recurrent neural networks (RNN) provide good results on such tasks. Recently, a Convolutional Neural Network (CNN) has been explored for detecting fake speech; in this direction, we propose a Convolutional Neural Network (CNN) model that inputs raw waveform after preprocessing and predicts whether the output is real or fake. The tested Convolutional Neural Network (CNN) model has given significant results comparable with Recurrent Neural Network (RNN) models.