Text independent automatic speaker recognition using selforganizing maps

This work presents one implementation of an automatic speaker recognition system, based on selforganizing map (SOM) neural networks. The voice of each speaker is modeled by a SOM, trained to specialize in the quantization of feature vectors (MFCCs) extracted from his voice. When a test sample is presented, it is quantized by all SpMs, that compete for the speaker: the SOM with smallest quantization error defines the speaker. The system was tested on a speaker identification task over a 14 speaker set, with phrases from three phonetically balanced sets and one variable answer set. The results comprovate the method's efficiency.