Using i-Vector Space Model for Emotion Recognition

Using i-vector space features has been shown to be very successful in speaker and language identification. In this paper, we evaluate using the i-vector framework for emotion recognition from speech. Instead of using standard i-vector features, we propose to use concatenated emotion specific i-vector features. For each emotion category, a GMM supervector is generated via adaptation of the neural one from a large corpus. An i-vector feature vector is then obtained using each emotion specific GMM supervector. The concatenation of these emotion dependent i-vector features is used as the feature vector in the SVM model for emotion classification. Our experimental results on acted and spontaneous data sets demonstrate that our proposed method outperforms other systems using either static features or GMM supervector features, and that system combination yields additional gain.