Connectionist Speaker Normalization with Generalized Resource Allocating Networks