Data-driven high-level information for text-independent speaker verification

Recently, various studies have shown that high-level features, such as linguistic content, pronunciation and idiolectal word usage, convey more speaker information and can be added to the low-level features in order to increase the robustness of the system. Usually these features are extracted by analyzing streams produced by phonetic speech recognition systems. Two of the major problems that arise when phone based systems are being developed are the possible mismatches between the development and evaluation data and the lack of transcribed databases. We propose in this paper to replace the phone-based approaches by data-driven segmentation methodologies. Our data-driven high-level systems do not use transcribed data and can easily be applied on development data minimizing the mismatches. These systems were fused with a state-of-the-art acoustic Gaussian Mixture Models (GMM) system. Results obtained on the NIST 2006 Speaker Recognition Evaluation data show that the data-driven features provide complementary information and the resulting fused system reduced the error rate in comparison to the GMM baseline system.