Intelligent system of speech recognition using Neural Networks based on DCT parametric models of low order

This paper proposes the development of a Numerical Command Recognition System of Speech Signal based on Neural Networks and DCT models. Thus, two configurations of neural networks, the Multilayer Perceptron and Learning Vector Quantization are evaluated by their performance in speech signal recognition, whose encoding is made by the mel-cepstral coefficients that are used to generate a two-dimensional time matrix by Discrete Cosine Transform (DCT). The selection of the best configuration of neural network for classification of the patterns was carried out by comparative analysis of performance of the MLP and LVQ networks through training, validation and test of the network topology and learning algorithms previously established. For demonstration of the performance of the proposed analysis methodology, the obtained results were compared with other methods of classification given by Gaussian Mixture Models (GMM) and Support Vector Machines (SVM).