Machine learning based speaker gender classification using transformed features

Speech and image processing are fundamental components of artificial intelligence technology. Speech processing can be deployed to acquire unique features of a person's voice. These can then be used for speaker identification as well as gender and age classification. This paper studies the effect of the relative degree of correlation in speech features on gender classification. To this end, gender classification performance is evaluated using orthogonally transformed speech features. The performance is then compared to the case when speech features are used without transformation. Two machine learning approaches are used in the evaluation. One of them primarily depends on Gaussian Mixture Models (GMM) and the other one uses Support Vector Machines (SVM). The results show that less correlated speech features, obtained after the orthogonal transformation, provides better classification performance.