Accuracy and Speed Performance Improvement in Speaker Verification Using Genetic Programming

In speaker verification, a system investigates a person's identity and decides whether the person is a true client or an imposter. In this paper, genetic programming (GP) is used as a method for speaker modeling. When GP is used for construction of models for speakers, due to long training time to train GP models, training data compression is proposed in this paper. This idea reduced training time for 20 times. Training of several GP trees as a speaker's model is another idea presented in this paper to improve the speaker verification performance. In this method, training data are separated to a few clusters. Then a GP tree is trained for each cluster. Therefore, a speaker is modeled by several genetic programming trees. The verification performance increased from 50% to about 92% using the proposed method. Genetic programming performance was compared to some other discriminative methods such as Multi-Layer Perceptron neural network and Learning Vector quantization, and generative methods such as K-Means, GMM and LBG, GMM-UBM and VQ-MAP. Experiments show that Genetic programming is more effective than the other methods.