Refining Cosine Distance Features for Robust Speaker Verification

Cosine distance features (CDF) have been proposed recently as a means to increase the performance of the speaker verification system. In CDF, we measure the cosine similarity of the i-vector of the input speech utterance with the i-vectors of the reference speakers. It is found that, the performance could be improved if the reference speakers are acoustically similar to the target speaker. There are different set of reference speakers for every target speakers if we select acoustically similar reference speakers, which is leading to speaker specific CDF(SSCDF). In this work, we explore the possibilities of further improving the performance of the SSCDF for the same number of reference speakers. We have developed two sub-systems with the reduced number of reference speakers (acoustically similar to the target speakers) and then combined/fused the decision scores of the two sub-systems. It is found that the fused SSCDF system out-performed the SSCDF consistently for reduced feature dimension and therefore using less number of reference speakers.