Partial Least Squares Based Total Variability Space Modeling for I-Vector Speaker Verification

As an effective and low-dimension representation for speech utterances with different lengths, i-vector method has drawn considerable attentions in speaker verification. Training a Total variability space (TVS) is one of the key parts in the i-vector method. However, the traditional training method only explores the relationship between different mean supervectors, ignoring priori category information of speakers, which results in a lack of discrimination. In the proposed method, a discriminative TVS based on Partial least squares (PLS) is estimated, in which both the correlation of intra-class and the distinction of inter-class are fully utilized due to using speaker labels, and the proposed method can achieve a better performance.