Language Effect on Speaker Gender Classification Using Deep Learning

In speech processing, identifying the speaker’s gender has been considered a topic of interest by many studies. Various approaches and methods have been proposed to detect the gender of a speaker with high accuracy. However, they are limited to isolated and specific languages. In this research, the speaker’s gender is classified from a mixed languages speech point of view, constituting six different languages using Bidirectional Long Short-Term Memory (BLSTM) network classifiers. Also, gender classification is performed using each specific language independently. The main aim of this approach is to tackle the effect of the language on speakers’ genders classification. Performance evaluation of the language effect on speaker gender classification is conducted on the open-source Mozilla datasets. We achieved an average gender classification accuracy of 90.42%, 97.42%, 82.44%, 98.39%, 100%, and 85.04% on Arabic, Chinese, English, French, Russian, and Spanish datasets, respectively. These results uncover some dependencies of speakers’ gender classification on the language.