Comparison of geometric, connections and structural techniques on a difficult isolated word recognition task

The sequential structure and variable length of speech data suggests the use of structural techniques such as Hidden Markov Models or Grammatical Inference systems. In contrast, geometric and classical (non-recurrent) connectionist methods deal with objects represented in a vector space. This means that some method has to be used to transform variable-length strings of parameters into d-dimensional vectors. Several such methods exist and some of them have been tested in this work in a difficult isolated word recognition task. The results of experiments with kNearest Neighbor, Multilayer Perceptron and Decision Surface Mapping are compared with other already reported using Hidden Markov Models, Error Correcting Grammatical Inference and Morphic Generator Grammatical