Speaker Gender Classification and Error Analysis Using Deep Learning and YOHO Corpus

Nowadays, digital speech processing has become a required field due to the evaluation of communication technology. Speech signals for any speaker hold meaningful information such as identity, gender, and age. Research into the gender and age of speakers has shown remarkable results in using deep learning architectures to capture the time and frequency pattern of the speech signal. Deep learning confers researchers the motivation to investigate speaker classification methods more deeply. In this work, we develop a speaker gender classification system using Bidirectional Long Short-Term Memory (BLSTM). The developed system is based on the prosodic features of the speech signal, such as speech fundamental frequency and log energy of the signal and their derivations. The considered prosodic features of the system give a novel method in the gender classification system. The system adapts using the public YOHO speech dataset and gives an overall accuracy of 98.59%, with 99.39% and 95.96% average accuracy for males and females, respectively. Also, human being perceptual classification is considered for critically missed speakers. This is helpful in order to uncover reasons for high errors for the purpose of increasing the robustness of the system regarding features selection and system tune up.