Improving channel robustness in text-independent speaker verification using adaptive virtual cohort models

In speaker verification, score normalization methods are a common practice to gain better performance and robustness. One kind of score normalization is cohort normalization, which uses information about the score behaviour of known impostors. During enrolment, impostor verifications are simulated to get a speaker-specific set of the most competitive impostors (the cohort). In the present paper, one virtual cohort speaker is synthesized using the most competitive impostor´s Hidden Markov Models (HMMs). These impostors are also users of the system and therefore their models have channel-specific information contrary to the universal background model, which provides channel- and speaker-independent models. On verification, cohort scores are obtained by an additional verification of the virtual cohort speaker. The cohort scores evaluate the candidate as an impostor. A cohort normalized score promises greater robustness. This paper will study the effect of the introduced cohort normalization technique on the speaker verification system atip VoxGuard, which is based on mel-frequency cepstral coefficients and HMMs. VoxGuard can be used as either a text-dependent or a text-independent verification system. In this paper, emphasis is placed on text-independent speaker verification. Experiments using the atip speech corpus and the SieTill speech corpus showed improvements measured by the equal error rate on performance and robustness.