Combination of self-organizing maps and multilayer perceptrons for speaker independent isolated word recognition

A new Neural Network architecture that combines the Kohonen Self-Organizing Maps and Multilayer Perceptrons for a speech recognition task is presented. This architecture overcomes the problem of time-alignement of the succesive frames obtained from one utterance of one word: the succesive frames of a word generate a trajectory in a two-dimensional space using the Self-Organizing Map. These are classified using the Perceptron. Comparation with other techniques are made, and results are better than the obtained wich Trace Segmentation. The vocabulary used in the experiments is a highly difficult subset from the Spanish alphabet: the Spanish E-Set.