Text Independent Speaker Recognition Using Fuzzy Based Neural Network

In the present scenario, the field of biometrics has arisen as an applied science with the goal of automatically measuring individuals' unique distinguishing traits and exploiting the data for forensic, security, and surveillance purposes. In text-independent speaker recognition, the speaker is recognized only on the basis of their voice signals, without reference to the content of the speech. Our solution to this is a fuzzy-based neural network [1], [15] architecture that learns feature vectors from training samples that contain MFCC. Three layers make up our suggested architecture: (i) the input layer, which accepts feature vectors as input; (ii) the hidden layer, which is in charge of employing k-means clustering to cluster the feature vectors; fuzzification; and computing effective information. and (iii) the output layer where the weighted information is converted into an objective function which is helpful for the unsupervised learning of the proposed network using the particle swarm optimization (PSO) [12]. The unknown parameters in the objective function to be optimized by PSO include the fuzzification parameters and the weights on the information computed at the hidden layer. In our work, we used a database with 100 speakers and 10 samples per speaker, and we created a novel architecture. 8 of these samples were used for testing, and the remaining 2 were used for training. After that, the trained network's performance was assessed using the test samples. In noise-free voice samples, our proposed network performed admirably with 87% accuracy, outperforming the Gaussian Mixture Model (GMM) at 86%.