Abstract
This paper describes a robust speaker dependent continuous digit recognition system which runs in real time on a 16-bit micro-processor. An important design goal was the efficient use of available processing resources. The decision-making steps are ordered according to the degree of difficulty and the amount of processing required. The system uses dynamic time alignment only selectively and locally, relying on lexical constraints imposed in the form of coarse phonetic transcription and a preclassification step which does not require costly time warping in pattern matching. The system achieved 96.5% string accuracy and 99.1% digit accuracy on 540 digit strings (average length of 4 digits) collected from six speakers (4 male, 2 female).

This publication has 11 references indexed in Scilit: