Speaker identification using radial basis functions

This paper describes a text-independent speaker identification system based on Radial Basis Functions (RBF) networks. Both text-dependent and text-independent speaker identification experiments have been conducted. The database contains 7 sentences and 10 digits spoken by 20 speakers over a period of 9 months. LPC-derived cepstrum coefficients are used as the speaker specific features. The results show that RBF networks offer fast learning speed and good generalization even in text-independent mode. Moreover, a robustness test has been carried out which demonstrates that RBF networks provide sufficient information to produce a `no match' decision in speaker identification applications.