Fast feature-based preclassification of segments in continuous digit recognition

This paper describes the preclassification part of a real time speaker dependent continuous digit recognition system. The system's main characteristic is a fast feature-based word hypothesizer employing a combination of knowledge-based methods and pattern matching. This provides for a more reasonable behaviour, avoiding unintuitive errors typically found in conventional pattern matching systems. The processing can be broken up into four components: on-line segmentation according to four coarse phonetic classes, feature based matching focusing on voiced segments, word candidate generation and pruning and finally, candidate-adaptive pattern matching for decision making. The digit accuracy of the preclassifier, which includes the first 3 steps was found to be 96% in experiments using a total of 540 digit strings with an average length of 4 digits, collected from six speakers (4 male, 2 female).