Toward a Natural Speech Understanding System.

摘要

Template matching speech recognition algorithms have proven effective in isolated-word speaker-dependent applications. Improvements may permit these algorithms to be used for a wider range of applications. This report describes several experiments on a template-matching system that are prerequisites for extending the system to multi-speaker applications. One set of experiments was performed on a two-speaker database, evaluating the learning algorithm and recognition performance with these two speakers. A second group of experiments evaluated system performance as a function of the amount of template-matching that is permitted. Next, several tests were conducted to evaluate system performance on a new database of eight speakers. In addition to the recognition tests, we identified three different acoustic representations. The three representations are successively closer approximations to selective characteristics of the human ear. Finally, two new recognition vocabularies, based on important linguistic considerations, are described. (RRH)