Investigation of Speaker-Independent Word Recognition Using Multiple Features, Decision Mechanisms, and Template Sets.
Abstract
A system was developed to investigate speaker-independent word recognition. The system incorporates multiple decision mechanisms, features, and template sets to recognize whole word, isolated speech inputs. The fundamental design concept extends from other successful recognition research efforts incorporating multiple feature sets. Categories of features combined in the system included formant tracks, wide-band spectrogram, zero-crossing rate, linear predictive coding (LPC) coefficients, LPC gain term, and time. A total of fifty-five processed features were derived from the feature categories. An advanced speech analysis tool called SPIRE provided the computational functions necessary for raw feature extraction. Preliminary results with limited vocabulary inputs and predefined template sets indicate the system can successfully recognize whole word inputs from a range of independent speakers. For certain decision mechanisms and template sets, recognition accuracy achieved better than 90% for first alternative word candidates. The system is implemented in the Lisp programming language on Symbolics 3600 series computers. (Thesis)
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1986
- Accession Number
- ADA178172
Entities
People
- Michael A. Brusuelas
Organizations
- Air Force Institute of Technology