No abstract

No abstract

No abstract

Resumo: O problema de reconhecimento automático de discurso tem se tornado alvo de estudos, em partes, devido à grande demanda por ferramentas que realizem esse processo de maneira rápida e eficaz. Este artigo investiga a aplicação da rede neural auto-organizável (SOM) de Kohonen para análise de fonemas da língua portuguesa em sinais de voz. Especificamente, a rede neural foi aplicada num difícil problema de reconhecimento de discurso contínuo, dependente do orador e com alfabeto aberto. Experimentos computacionais revelaram que a SOM é capaz de identificar corretamente muitos fonemas de um discurso contínuo. Porém, a rede também identificou alguns fonemas inoportunos que podem estar relacionados à transição de palavras, rotulação manual errada, ou que possuem sons semelhantes.Palavras-chave: mapas auto-organizáveis; reconhecimento automático de discurso; redes neurais artificiais.

This dissertation studies a class of evolving neural networks for system modeling from data streams. The class encompasses single hidden layer feedforward neural networks with variable and online definition of the number of hidden neurons.Evolving neural network learning uses clustering methods to estimate the number of hidden neurons simultaneously with extreme learning algorithms to compute the weights of the hidden and output layers. A particular case is when the evolving network keeps the number of hidden neurons fixed. In this case, the number of hidden neurons is found a priori, and the hidden and output layer weights updated as data are input. Clustering and extreme learning algorithms are recursive. Therefore, the learning process may occur online or real-time using data stream as input.Two evolving neural networks are suggested in this dissertation. The first is an evolving hybrid fuzzy neural network with unineurons in the hidden layer. Unineurons are fuzzy neurons whose synaptic processing is performed using uninorms. The output neurons are sigmoidals. A recursive clustering algorithm based on density and data clouds is used to granulate the input-output space, and to estimate the number of hidden neurons of the network. Each cloud corresponds to a hidden neuron. The weights of the hybrid fuzzy neural network are found using the extreme learning machine and the weighted recursive least squares algorithm. The second network is an evolving multilayer neural network with sigmoidal hidden and output neurons. Like the hybrid neural fuzzy network, clouds granulate the input-output space and gives the number of hidden neurons. The algorithm to compute the network weights is the same recursive version of the extreme learning machine. A variation of the adaptive OS-ELM (online sequential extreme learning machine) network is also suggested.Similarly as the original, the new OS-ELM fixes the number of hidden neurons, but uses sigmoidal instead of linear neurons in the output layer. The new OS-ELM also uses weighted recursive least square. ix The hybrid and neural networks were evaluated using two classic benchmarks: the gas furnace identification using the Box-Jenkins data, and forecasting of the chaotic Mackey-Glass time series. Synthetic data were produced to evaluate the neural networks when modeling systems with concept drift and concept shift. This a modeling circumstance in which system structure and parameters change simultaneously.Evaluation was done using the root mean square error and the Deibold-Mariano statistical test. The performance of the evolving and adaptive neural networks was compared against neural network with extreme learning, and evolving modeling methods representative of the current state of the art. The results show that the evolving neural networks and the adaptive network suggested in this dissertation are competitive and have similar or superior performance than the evolving approaches proposed in the literature.