ESCA Workshop on Automatic Speaker Recognition, Identification, and Verification

Martigny, Switzerland
April 7-9, 1994

Text-Independent Speaker Identification with Functional-Link Neural Networks

Pierre Castellano, Sridha Sridharan

Signal Processing Centre, Queensland Institute of Technology, Brisbane, Australia

The discriminatory capabilities of Speaker Identification (SI) neural networks have traditionally been attributed to the presence of one or two hidden layers in their architectures. This study compares the speaker identification performance of supervised flat functional-link neural networks firstly to that of a multi-layered back-propagation network and secondly to that of the Interactive Laboratory System. In this study and given a 14 speaker database, weights in the layered model were unable to converge to global minimum. This was not the case for the flat architectures. Tensor based functional-link flat networks were found to be superior ( 73 percent correct speech frame classification) to hidden layer architectures ( 30-40 percent correct frame classification). The flat network's SI performance was comparable to that of an established SI technology given a 20 speaker database independent of the first. However the network was three times slower than the latter.

Full Paper

Bibliographic reference.  Castellano, Pierre / Sridharan, Sridha (1994): "Text-independent speaker identification with functional-link neural networks", In ASRIV-1994, 111-114.