Unwarping of continuous speech signals

Unwarping of varying continuous speech patterns with neural arrays is studied. These curved arrays are designed to solve the nonlinear time-alignment problem. We apply the self-organization algorithm to drive and stretch these flexible arrays in a 2-D plane to fit the speech patterns. These arrays plus their stretching parameters constitute a set of time-scaled reference templates. With these templates, the matching between unknown speech patterns and templates can be evaluated accurately. Two perception energies are devised to count the matching credits. One energy counts the time precedent relationships among the patterns. The other energy counts the matched features without considering the precedent relationships. A speech system which can recognize finite Chinese words is developed based on the proposed approach