ISCA Archive Eurospeech 1993
ISCA Archive Eurospeech 1993

Modelling speaker normalization by adapting the BIAS in a neural net

David J. M. Weenink, Louis C. W. Pols

We present a possible model for speaker normalization in terms of adaptation. Vowel identification is considered to be based on the integration of a number of decisions. Each decision is of the same simple form: decide whenever the accumulated evidence exceeds a certain threshold (bias). Changing the bias can influence a decision and consequently the identification. This process just as well describes human vowel perception as neural net simulation.

Keywords: Speaker normalization, neural nets