Speech analysis and recognition using a feature-based weighted scoring mechanism

A simple scoring mechanism has been developed for speech recognition purposes. In analysis mode (common to both training and recognition processing), peak patterns are obtained from each 10 ms frame of channel values from a 15-channel analog filter bank. In addition, tests are made for the presence of certain high-level speech features such as nasality, unvoiced speech etc. Scoring is based on matching input peak patterns and features against stored patterns and features. Successful matching results in the score associated with a particular peak pattern or feature being added to an accumulated score. Weightings are used to discriminate between features and peak patterns and between peak patterns