Reconocimiento de patrones vocálicos mediante la implementación de una red neuronal artificial utilizando sistemas embebidos

Abstract

In this work the results of the design and development of a neural network type feed-forward in an embedded system, to identify patterns associated with the pronunciation of vowels of Spanish language are presented. For the training of the network Mel frequencies cepstrals coefficients extracted from audio signals related to the pronunciation of the open vowels (/ a /, /e/ and /o/) were used. The capture of new samples was performed by using a computer and an embedded system to evaluate, test and compare the performance of the neural network. A percentage of correct identification above 98% was obtained indicating that the error in the class separation was less than 2% for patterns obtained from any of the three vowels. Based on the results it is concluded that implementation of artificial intelligence algorithms for classification tasks in embedded systems is feasible, and presents similar results to those that would have the same system working with the resources of a computer.