Fast algorithms to find invariant features for a word recognizing neural net

A short description of the feature finding neural net (FFNN) for the recognition of isolated words will be given. As has been shown elsewhere, during recognition mode FFNN is faster than the classical HMM and DTW recognizers and yields similar recognition rates. In this article the emphasis is placed on optimal and fast algorithms for selecting relevant features from the speech signal. By the growth algorithm it is possible to increase the network's size gradually by adding relevant feature detecting cells. The substitution algorithm starts with a full-size net and arbitrary features. Then it substitutes less relevant features by features with higher relevance. Recognition results for both cases will be given and discussed. Finally, it will be shown that with FFNN it is possible to solve a special case of the figure-ground problem.<
>