Spoken emotion recognition via locality-constrained kernel sparse representation