Speaker Recognition with Mismatched Coded Speech

This paper investigates the effects of low-bit rate coded speech on the performance of a fixedtext speaker recognition system, under mismatched coding conditions between enrollment and testing. Significant degradation of performance has been observed relative to matched conditions, where same coding is used. Two techniques have been proposed to overcome mismatch effects; a linear discriminative mapping of the mismatched data to the baseline models, and a neural network-based, reduced dimension feature extraction robust to the mismatched coding conditions. Key-Words:Fixed-text speaker recognition, speech coding, mismatched conditions, linear transform, neural networks, feature extraction.