Score normalisation applied to open-set, text-independent speaker identification

This paper presents an investigation into the relative effectiveness of various score normalisation methods for open-set, textindependent speaker identification. The paper describes the need for score normalisation in this case, and provides a detailed theoretical and experimental analysis of the methods that can be used for this purpose. The experimental investigations are based on the use of speech material drawn from 9 hours of recordings of different Broadcast News. The results clearly demonstrate the significance of improvement offered by score normalisation. It is shown that, amongst various normalisation methods considered, the unconstrained cohort normalisation method achieves the best performance in terms of reducing the errors associated with the open-set nature of the process. Furthermore, it is demonstrated that both the cohort and world model methods can offer very similar effectiveness, and also outperform the T-norm method in this particular case of speaker recognition.