Trabalho de Graduação Interdisciplinar apresentado ao Curso de Engenharia Elétrica, da Escola de Engenharia da Universidade Presbiteriana Mackenzie, como requisito parcial à obtenção do grau de Bacharel em Engenharia.
Este trabalho apresenta os conceitos básicos e as técnicas envolvidas na implementação de um sistema de reconhecimento de voz via Redes Neurais Artificiais, utilizando o programa Matlab®.
Inicialmente, foi realizado um estudo dos fatores que dificultam o reconhecimento de voz, tais como: formação do aparelho fonador, estilo de falar e sotaque.
Posteriormente, foram apresentadas as formas de pré-processamento e algumas técnicas de extração de características dos sinais de voz, essenciais no processo de treinamento de um classificador neural.
Em seguida, foi apresentada uma descrição detalhada da arquitetura da Rede Neural Perceptron Multicamada adotada como classificador.
Por fim, os resultados experimentais são descritos. Estes demonstram a viabilidade de uma rede neural como classificador em um sistema de reconhecimento de voz.
Palavras-chave: Reconhecimento de Voz, Rede Neural Perceptron Multicamda, extração de características, classificador, pré-processamento.
This are my paper presented to the Engineering Course Electrical, from the School of Engineering at Mackenzie Presbiterian University, as a partial requirement for obtaining the degree of Bachelor of Engineering.
This work presents the basic concepts and techniques involved in the implementation of a speech recognition via Neural Networks using Matlab® software.
Initially, it was accomplished a study of the main factors that hinder the speech recognition system, such as: the constitution of the articulatory apparatus, style of speaking and accent.
Later, it were presented preprocessing method and the feature extraction techniques for speech recognition, essential elements in the process of training of a neural classifier. After that, a detailed description of the architecture of Multi-Layer Perceptron Neural Network, adopted as classifier, was presented.
Finally, the experimental results were described. These demonstrate the viability of a Neural classifier in a speech recognition system.
Key-words: Speech Recognition System, Multi-Layer Perceptron, Neural Network, feature extraction techniques, preprocessing method.