A combined approach to speaker authentication using claimant-specific acoustic universal structures

In this paper, a novel approach to online cohort selection is proposed which combines the cohort sets obtained using acoustic universal structure (AUS) and speaker specific cohort selection (SSCS). To obtain the cohort set using AUS, a confusion matrix is first generated using the distance between the structure of test utterance and the AUS of speaker. The confusion matrix is normalized using the iterative proportional fitting (IPF) method. The normalized confusion matrix along with a simple distance metric is used to select a cohort set based on similarity for each client speaker. A similar procedure is followed to obtain the cohort set using the SSCS method. Both these cohort sets are combined to obtain a single cohort set. Normalization statistics are then computed from this cohort set, which is used in the final scoring for authenticating the claimed speaker identity. Speaker verification experiments conducted on the NIST 2002 SRE, NIST 2004 SRE and YORO database, show reasonable improvement over conventional techniques used in speaker verification in terms of equal error rate and decision cost function values.