Speaker identification using hybrid LVQ-SLP networks

The architecture and learning strategy of a hybrid LVQ-SLP (learning vector quantization and single-layer perceptron) network aimed at speaker identification are introduced. Its performance is compared with two of the most popular networks: LVQ and MLP networks. The hybrid LVQ-SLP network is characterized by the following properties: (1) it makes use of the existing training algorithms developed for LVQ and MLP networks; (2) it provides identification performance comparable to that of our best MLP network but with less training time and considerably outperforms the performance of the corresponding LVQ network. In a text-independent speaker identification experiment with 112 male speakers, the identification rate by the hybrid LVQ-SLP network is 97.3%, while the corresponding LVQ network with the same codebook gives only 83.5%.