Design of hierarchical perceptron structures and their application to the task of isolated-word recognition

Abstract

Several design strategies for feed-forward networks are examined within the scope o f pattern classification. Singleand two-layer perceptron models are adapted for experiments in isolated-word recognition. Direct (one-step) classification as well as several hierarchical (two-step) schemes have been considered. For a vocabulary o f twenty English words spoken repeatedly by eleven speakers, the word classes are found to be separable by hyperplanes in the chosen feature space. Since for speaker-dependent word recognition the underlying data base contains only a small training set, an automatic expansion o f the training material improves the eneralization properties o f the networks. This methoc?accounh for a wide variety o f observable temporal structures for each word and gives a better overall estimate of the network parameters which leads to a recognition rate o f 99.5 %. For speaker-independent word reco nition, a hierarchical structure with pairwise training of?wo-class models is superior to a single uniform network (98 % average recognition rate).

A three-hierarchy neural network structure and its application to the speaker-independent speech recognition of Chinese words

“…It is known that speech features are easy to confuse, however this means of painvise class discrimination is fit for distinguishing easy-to-confuse classes owing to its some fuzzy properties inherently. Therefore the English speech recognition by the approach in [2] has reached a good effect.…”

“…Moreover, the speech spectrums of some different digits are very similar to each other, sich as 2 and 8. So the effect is not very good when the peaker-independent Chinese speech recognition is implemented by the approach adopted in [2]. Therefore many experiments are carried out on several distinct hierarchic structures, and large numbers of spectrums of speech templates are analysed carefully.…”

“…This three-hierarchic structure mainly adds the first or sorting network by comparison with [ 2 ] . The sorting network separates the classes from the other ones according to the marked difference of spectrums, and gathers up the classes whose spectrums are relatively similar into a sort.…”

“…Paper [2] presents a neural network of twohierarchic structure for the speech recognition of a small vocabulary of English words by means of painvise class discrimination. It is known that speech features are easy to confuse, however this means of painvise class discrimination is fit for distinguishing easy-to-confuse classes owing to its some fuzzy properties inherently.…”

“…Based on [2], this paper proposes a three-hierarchic structure of neural networks through the agency of many experiments and a good deal of research. Furthermore, a speaker-independent speech recognition system for Chinese digits 0-9 is developed with the structure.…”