Comparing various automatic speaker recognition approaches [articol]

Abstract

Connectionist methods in speaker recognition give promising results on the condition of selecting a well balanced feature vector. Special attention has to be paid to the classifier procedure applied on the presented templates. We are discussing here several different speaker recognition methods, focusing on two main approaches: a MFCC supervised approach using a Hausdorff-based metric and an AR - ART2 structure (autoregressive features classified with an Adaptive Resonance Theory type 2 Network). Then, we report them to our previous results. These methods can function in a single or parallel structure, needing a weighted inference to accomplish the recognition task. Respecting the well known a-priories in signal treatment, and carefully selecting the frequency bands to be “examined”, our trials gave good results, with a mean detection error reduced with about 3% compared to our previous essays. Our studies concentrate on tests made with discrete-words, in real, noisy environment and real-time / off-line speaker verification.