The relation of acoustic phonetic states to the problem of speech recognition is discussed. It is suggested that a first step toward understanding the dynamics of the speech signal could be made by recognizing small number of acoustic‐phonetic classes. Adaptive threshold elements are proposed as a means of recognizing these classes and a method of utilizing the adaptive threshold elements in a decision procedure is presented. The means employed to gather the data representing the classes consists essentially of taking amplitude samples from a bank of 15 filters at 10‐msec intervals. Samples obtained this manner are used as input for the adaptive‐decision procedure, which is simulated in a general‐purpose computer. Samples representative of each class are used to train the adaptive‐decision procedure, and the capability to generalize to new samples is observed. For one speaker, generalization results of 92%‐correct sample classification were achieved, and generalization from one speaker to another was demonstrated. It further was shown that the selection of an output code can significantly affect the generalization and that sequences of recognized samples can represent dynamic changes through words.

This content is only available via PDF.
You do not currently have access to this content.