MULTIDIMENSIONAL MODEL FOR AUTOMATIC SPEECH RECOGNITION
Abstract
A theoretical basis is provided for a general purpose speech recognizer. The research has focused upon the nature of normal speech, which can be distinguished from discrete articulation by the continuous movement (in normal speech) of articulators from one position to another; as a result, sounds in continuous speech are more likely to modify the production of surrounding sounds than they are in discrete speech. Assuming that, according to the ergodic theory, sound changes occurring in everyday speech reflect and repeat the changes occurred in the historical development of language (because the physical modes of speech production are the same), linguistic examples and theories of sound change were studied. From this study, a body of rules for sound change or euphonic combination was derived and their applicability to the English language tested. These rules represent an error-correcting code to restore omitted or indefinite word boundaries and/or to restore the orthographic phone classes which are altered in continuous speech. The study required the evaluation of existing research and theories, as well as the generation of some original data, the latter consisting of high-quality recordings of continuous speech samples. Both original data and previously published data were subjected to acoustic analysis of minute portions of the speech waveform.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 14, 1964
- Accession Number
- AD0437324