Gender classification by processing emotional speech

Gender information improves emotion classification accura cy. Accordingly, gender classification for emotional speech is an interesting research t opic. A pool of 1418 features is created, 619 of which are tested for the first time in gender cl assification. This is the largest feature set to the best of the authors’ knowledge. The fe atures are related to statistics of pitch, formant, and energy contours, as well as autocorrelation, MPEG-7 descriptors, Fujisaki’s model parameters, jitter, and shimmer. A featur e selection algorithm derives 15 features, which are input to a nearest neighbor classifier, radial-basis function neural network, probabilistic neural network, support vector machines, discriminant analysis-based classifier, classification tree, self organizing map, and ne ural gas network. Two databases are employed: the Berlin database of Emotional Speech and the Danish Emotional Speech database. A perfect classification accuracy is obtained. A s ystematic comparative study of the performance gains among the classifiers and the variants of each particular classifier is undertaken. Furthermore, the databases are assessed concerning gender classification accuracy. Results advance the state-of-the-art.