Frame-level features conveying phonetic information for language and speaker recognition

Abstract

This Thesis, developed in the Software Technologies Working Group of the Departmentof Electricity and Electronics of the University of the Basque Country, focuseson the research eld of spoken language and speaker recognition technologies.More specically, the research carried out studies the design of a set of featuresconveying spectral acoustic and phonotactic information, searches for the optimalfeature extraction parameters, and analyses the integration and usage of the featuresin language recognition systems, and the complementarity of these approacheswith regard to state-of-the-art systems. The study reveals that systems trained onthe proposed set of features, denoted as Phone Log-Likelihood Ratios (PLLRs), arehighly competitive, outperforming in several benchmarks other state-of-the-art systems.Moreover, PLLR-based systems also provide complementary information withregard to other phonotactic and acoustic approaches, which makes them suitable infusions to improve the overall performance of spoken language recognition systems.The usage of this features is also studied in speaker recognition tasks. In this context,the results attained by the approaches based on PLLR features are not as remarkableas the ones of systems based on standard acoustic features, but they still providecomplementary information that can be used to enhance the overall performance ofthe speaker recognition systems.