Interfusing the Confused Region Score of Speaker Verification Systems

In the text-independent speaker recognition field, there have been many excellent techniques based on the cepstral acoustic features. Recently, high level prosodic features have been widely used to verify the speaker's identity as they are less sensitive to the channel and noisy effect. But how to combine the existed prosodic system's scores with the scores of the system based on acoustic features to achieve a superior performance becomes a very difficult issue. This paper presents a combination method called interfusing the confused region scores (ICRS) to achieve a better recognition precisio.n. We report results on the NIST 2006 speaker recognition evaluation (SRE) using two component systems: a standard MFCC-SVM and PGCP- SVM prosodic system, and show that the proposed interfusing technique results in 9.25% reduction in equal error rate (EER) and at last gets a EER = 4.9% after system's combination.