Automatic Gender Identification Using Fusion of Generative and Discriminative Classifiers and Clustering of Spekaers from the Same Gender

This paper proposes a two layer classifier fusion technique using clustering of training data from speakers of the same gender for automatic gender identification (AGI). The first layer is an acoustic classification layer for mapping MFCC and pitch acoustic feature space to score space. In this layer, a divisive clustering is proposed for dividing the speakers from each gender to some classes, where speakers in each class have similar vocal articulatory characteristics. Finally, the best structure could map 22 feature coefficients to 5 likelihood scores as new features. The second layer is a back-end classifier that receives the vectors of fused likelihood scores from the first layer. This means that the new feature coefficients are used in the second layer. GMM, SVM and MLP classifiers were evaluated in the middle and back-end layers. 96.53% gender classification accuracy was obtained on OGI multilingual corpus which is much better than the performance obtained by traditional AGI methods.