Frame Length Selection in Speaker Verification Task

In this paper an approach based on the use of different frame lengths for the feature extraction process in the training and recognition phases of a speaker verification system is presented. It is able to sensibly reduce the ER and the degradation on performance related to the training/verification feature mismatch. The potentiality of the approach are investigated in an a-posteriori search of the best combination to be adopted. A prototype of an expert system is also presented in order to automatically search, in real time operating conditions, the optimal combination of parameters. Tests have been performed on a set of speakers whose speech productions were spanned over approximately 3 months. Key-Words: Speaker Verification, Text Dependent, Mismatch, Frame Length, CD-HMM