Modular networks and constraint satisfaction model for recognition of stop consonant-vowel (SCV) utterances

The main issues in the recognition of stop consonant-vowel (SCV) utterances of Indian languages are the large number of classes and the high confusability among several classes. In modular approaches separate neural networks (subnets) are trained for subsets of classes, and the outputs of the subnets are processed to perform the classification. Performance of the modular networks is poor because the class of the largest value among the outputs of the subnets is assigned to the input. We propose a constraint satisfaction model (CSM) in which the outputs of the subnets are combined using the constraints that represent the similarities among the SCV classes. Results of the recognition studies show that the CSM gives significantly better performance (about 65%) compared to the conventional modular networks (about 35%)