Improved gender/age recognition system using arousal-selection and feature-selection schemes

摘要

This work proposes the arousal-selection and feature-selection schemes to improve speaker's gender and age identification performance. Our previous results showed that gender and age recognition rates would increase as affective stimulation degrees were lower and higher, respectively. Considering a practical scenario, the speaker's mood does not alter frequently, so speech frames are partitioned into two groups with low and high arousal levels. Here, two Gaussian Mixture Model (GMM) probability density functions are employed to characterize the distributions of the degrees of speech stimuli in terms of tone and energy variations. Such approach can appropriately classify speech frames and easily adapt to different speakers. As well as speech frames are fairly filtered and partitioned, the feature-selection scheme is effectively used to determine adequate low-level features. To do fair comparison, the experiment database adopts Lwazi corpus from South Africa. The proposed system using the arousal-selection and feature-selection schemes exhibits that accuracy rates of gender and age estimations reach 98.9% and 71.6% with 1.7% and 10.8% increases, respectively, as compared to the ones without using arousal-selection and feature-selection schemes. Therefore, the recognition system proposed herein successfully enhances accuracy rates of age and gender estimations for various human-machine interaction and multimedia applications.