Speaker identification based on complete corpus and multi-step mini-max search matching

A new text-independent speaker identification approach is described. It models speaker by complete feature corpus (CFC) trained from selected speech sample data with various and sufficient phonemes and phonetic phenomenon. The matching between input speech and the CFC models is executed with a multi-step mini-max search (MMS) algorithm, which evaluates the mutual information between the input speech and CFC models. Experiments on performance evaluation show that the proposed CFC model and MMS algorithm are quite effective.