Gender Classification of Mixing and De-mixing Speech

Abstract

Gender classification is growing in popularity due to the variety of fields in which it can be used. It can be employed in various fields, including criminal investigations and security and authentication services. Gender Classifying speech for different speakers is still a demanding and challenging task for recognizing overlapped speech and building a robust prediction model. The paper provides a gender classification system that uses Independent Component Analysis (ICA) and several machine learning algorithms to identify mixing and de-mixing speech signals. ICA is employed to separate the mixed signal into their source signals. The system consists of two stages: the first stage is the mixing and separating process for signals. The second stage involves combining feature extraction and constructing a classification model to determine whether a signal is male or female based on its acoustic attributes. The system will evaluate the efficacy and significance of machine learning algorithms for selecting the optimal method to identify the speaker's gender with the most excellent efficiency and accuracy. Experimentation shows that the best accuracy value for an SVM model with mixing speeches is 87.1 %, and the best accuracy value for a Neural Net and SVM model with de-mixing speeches is 97.8 %.