We compare several neural networks architectures to measure the degree of similarity among speakers. For each speaker of a reference set, Multilayer Perceptrons and Radial Basis Functions are trained to perform a non-linear principal component analysis of acoustic vectors, and Self-Organized Feature Maps are used to construct Vector Quantizers. As a first simple step, we use non-discriminant training to characterize speakers, and, then, the result is applied to combine speaker-dependent speech recognition models. In a second phase, discriminant training over speaker models is carried out, and speaker verification and identification performances of these networks are evaluated.

Keywords: Speech recognition, speaker recognition, neural networks, similarity measures, models.