Speech representation based on tensor factor analysis and its application to speaker recognition and language identification
Daisuke Saito, So Suzuki, Nobuaki Minematsu · 2019
Ahstract-This paper proposes a novel approach to speech representation for both speaker recognition and language identification by characterizing the entire feature space by a tensor. In conventional studies of both tasks, i-vector is commonly used as the state-of-the-art representation. Here, i-vector extraction can be regarded as projection of utterance-based GMM supervector onto a low-dimensional space. In this paper, for the aim of explicit modeling of the correlation among mean vectors of a GMM, an utterance is not modeled as its GMM-based supervector but as its matrix and the entire set of utterances is modeled as its tensor. By applying tensor factor analysis, we obtain a new representation for an input utterance. Experimental evaluations for speaker recognition and language identification show that our proposed approach has effectiveness especially for the speaker recognition task.