Transition-oriented Hidden Markov Models for Speaker Verification

摘要

In this article, we present a novel mechanism by which more precise voiceprints can be constructed in a typical text-dependent speaker verification system based on a continuous density hidden Markov mdoel (HMM)> Typical voiceprints (speaker-dependent HMMs) are first trained using a subscriber's enrollment data. The resulting models are then restructured to permit a modeling of sub-state behavior. At first, the restructured models are function-ally equivalent to the conventional voiceprint. Sub-state parameters are then estiamted by the re-application of the enrollment data. The resulting speaker-dependent models provide improved speaker verification performance relative to the models with the original topology.