INCREASING ROBUSTNESS TO TRAINING-TEST MISMATCH IN SPEAKER VERIFICATION THROUGH SHALLOWSOURCE MODELLING

摘要

Speaker verification is usually performed by comparing the likelihood score of the target speaker model to the likelihood score of an universal background model (UBM), and then applying a suitable threshold. For the UBM to be effective, it must be estimated from a large number of speakers. However, it is not always possible to gather enough data to estimate a robust UBM, and the verification performance may degrade if impostors, or whatever sources that generate the input signals, were not suitably modelled by the UBM. In this work, a new normalization technique is proposed, based on a shallow source model (SSM) estimated from the input utterance. A linear combination of the likelihood scores of the SSM and the UBM is used to normalize the speaker score. Speaker verification experiments were carried out on a clean-speech dataset including 204 speakers. Also, a sizeable amount of noisy, speech and non-speech signals was used to test the robustness to large training-test mismatch. Three normalization techniques were tested: UBM, smoothed UBM and the proposed combination of UBM and SSM. This latter approach yielded the best performance. The difference in performance was specially significant in the large training-test mismatch condition.