Voice-based gender identification via multiresolution frame classification of spectro-temporal maps

This paper presents a novel approach to gender identification based on adaptive multiresolution (MR) classification of spectro-temporal maps. The images of speech signals in this work are mainly provided by auditory inspired spectro-temporal representations: mel-spectrogram, cochleagram and auditory spectrogram. The 2-D representation of a segment of an utterance is used as the input to the system. The system adds MR decomposition in front of a generic classifier consisting of feature extraction and classification in each MR subspace, finally combined into a global decision using a weighting algorithm. It has been shown that the accuracy of the proposed method, by rising up to 99%, significantly outperforms the accuracy of most of other common algorithms which combine pitch and acoustical features for gender identification.