Neural network architectures for speech recognition

Artificial neural networks (NNs) are a popular approach in the area of speech recognition, but several problems still exist to fulfil the proposed tasks, such as type of architecture, number of layers and cells, and how to deal with training processing time. The paper describes the results obtained from an experimental speech recognition system designed to compare several NN architectures in the speech recognition task. To perform this research some experiments have been undertaken using different groups of data and several speech features. These experiments investigate the performance of several NN architectures with different number of layers, different number of cells and different learning algorithms in order to deal with processing time and the local minima problem.<
>