Comprehensive Analysis of Deep Learning Methods on Gender Identification Using Speech Signals

Human-Computer Interaction Systems have been becoming ubiquitous owing to ever-extending information generation. However, the rising numbers also breed credibility questions about these systems. As one of the fundamental information carried by speech signals, the gender identity of the person bears the potential to solve many authentication problems. In this study, we surveyed several Convolutional Neural Network architectures in the binary gender identification task. As our experimental results demonstrated, the VGG network showed superiority over the other model by achieving 86.01% weighted F1 in the CommonVoiceTR-9.0 Turkish dataset and 88.32% weighted F1 in the IEMOCAP English dataset. To the best of our knowledge, this is the first study surveying deep learning methods for the speech gender identification task in Turkish. Moreover, the capabilities of the models in language independency were questioned in this study.