ADEQUACY OF NEURAL PREDICTORS FOR SPEAKER IDENTIFICATION

We analyse in this paper neural prediction systems for automatic speaker identification (A.S.I.) and links between models complexity and their performances. We develop two ideas for enhancing such systems : we first reconsider the validity of standard hypothesis underlying the use of predictive models, we then propose different techniques for incorporating a-priori knowledge in our models. We illustrate the different points by providing results on 15 talkers from the TIMIT database.