Integration of MKL-based and I-vector-based speaker verification by short utterances

We developed a speaker verification system that is efficient for short utterances. The i-vector-based speaker representation has helped realize highly accurate speaker verification systems, however, it might be not robust against short utterances because the reliability of statistics required for extracting i-vectors is low. On the other hand, multiple kernel learning based on conditional entropy minimization has also achieved high accuracy in speaker verification that is robust against intra-speaker variability. To improve the robustness of speaker verification systems against short utterances, we attempted to integrate the above-mentioned complementary systems. Our experimental results showed that the proposed system integration achieved high-accuracy speaker verification systems, irrespective of the utterance lengths, even for very short utterances (e.g., less than two seconds).