Group sparse features for speech emotion perception in tensor space

With increasing demands for a natural interaction between human and machine, emotion perception from speech signals is becoming an important interaction interface. In this paper, we give a feature extraction framework for speech emotion recognition and present a novel method to extract emotion information based on group sparsity in tensor space. The speech signal is encoded as cortical representation in auditory system. We propose the group lasso nonnegative tensor factorization model to learn the multilinear factor matrices from tensor feature subspaces. l₁/l₂ constraint on multiple subspaces is imposed to recover the different groups of covariance for each factor (frequency, time, etc). The experimental results show that the proposed method can improve the multi-classes emotion recognition performance than state of the art baseline systems.