Fake Speech Recognition Using Deep Learning

The increase in the number of algorithms and commercial tools for creating synthetic audio has led to a high level of misinformation, especially on social media. As a consequence, efforts have been focused in recent years on detecting this type of content. However, this task is far from being successfully addressed, as the naturalness of fake audios is increasing. In this paper we present a model to classify audios between natural and fake, using an audio preparation stage that includes raw audio transformation, and a modelling stage by means of a custom Convolutional Neural Network (CNN) architecture. Our model is trained on data from the FoR dataset, which contains natural and synthetic audios obtained from several algorithms for deepfake content generation. The performance of the model is evaluated with different metrics such as F1 score, precision (P) and recall (R). According to the results, the audios are successfully classified in 88.9% of the cases.