Classifying Gender of Separated Voices From Independent Component Analysis Using Hybrid features
Sawsan Hadi Abed, Nidaa AbdulMohsin Abbas · 2022
Building a robust prediction model for classifying gender based on speech signals belonging to various speakers is still a complex computational and challenging process for identifying gender in overlapping or mixing signals. This paper presents a classification model for characterizing a speaker's gender by separated speech signals resulting from the separation of mixing signals using Independent Component Analysis. The gender of separated signals determines based on a set of features and using various Machin Learning algorithms. The adopted methodology in this paper involved two-fold, the first fold contains the mixing and demixing using fastica. While the second fold uses the hybrid feature extraction based on traditional features (MFCC, pitch, shimmer, formant) with classification methods represented by SVM, DT, ANN and RF. The performance results of the hybrid and traditional features are compared to find the optimum set of features for enhancing the classification models based on the considered voices database. The experimental results reveal the significance of hybrid features and their importance in improving the performance of the classification model.