Automatic Recognition of Spoken English Words.
Abstract
The object of the research was the development and testing of techniques for use in the recognition of spoken words. A trainable pattern classifier to recognize a limited vocabulary of spoken words was implemented using these techniques. The words used were spoken into a microphone located on the operator console in the Hybrid Computing Facility. Speech samples were processed by a 31-channel bandpass filterbank before digitization. Training was accomplished by first segmenting each spoken word into three time sections and then performing clustering operations on these segments to derive a set of prototype segments, along with estimated a priori probabilities of occurrence for these prototypes. The system was trained first on a single speaker and later on a group of five speakers, using the digits/zero/through/nine/as a test vocabulary. Recognition rates were generally from 94 to 98% on new data by the speakers for which the system was trained. (Author Modified Abstract)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 28, 1972
- Accession Number
- AD0757492
Entities
People
- C. L. Coates
- Gary Lynn Hunt
Organizations
- University of Texas at Austin