MATLAB Implementation of Speaker Recognition with Multiple Algorithms

This MATLAB-implemented speaker recognition program employs various algorithms including BP (Backpropagation), PNN (Probabilistic Neural Network), SOM (Self-Organizing Map), RBF (Radial Basis Function), and LVQ (Learning Vector Quantization) for speech file training and testing, achieving excellent recognition results. The bprengong program implementation involves two main phases: training and testing. The first phase focuses on computing the recognition model. The variable 'v' represents the feature vectors after MFCC (Mel-Frequency Cepstral Coefficients) processing. To handle variable-length speech data, the implementation performs uniform truncation to normalize the input dimensions. The matrix 'p' contains rows representing individual speech data samples (15 total samples), while 'Pr' stores the maximum and minimum values for each row for normalization purposes. The target variable 'T' contains the expected output values. The neural network architecture features 15 output neurons corresponding to the number of speaker classes. During training, when processing input samples with class label 'i', the implementation sets the expected output of the i-th neuron to 1 while all other neurons are set to 0. This one-hot encoding approach effectively trains the network for multi-class classification. During the recognition phase, when an unknown speech sample is presented to the network, the algorithm examines the output values from all neurons and assigns the sample to the class represented by the neuron with the highest output value. This max-output selection strategy provides robust speaker identification capabilities across different algorithmic approaches implemented in the program.