A Cost Constrained Speech Recognition Systems.
Abstract
This paper describes the design, construction, and evaluation of a microprocessor-based, cost-constrained word recognition system. The system uses seven bandpass filters, logarithmically spaced, followed by envelope detectors. The final algorithm uses eight uniformly spaced time slices, and used dynamic programming for time warping, with a weighted Tchebycheff distance. This system resulted in 98% correct recognition for the ten digits, 0 - 9, of the training group, and 96% correct recognition for the control group. The project demonstrated the necessity for an improvement gained with time warping. Rabiner's Unconstrained Endpoint Local Minima algorithm was used to perform the time warping. For the system used, it was found that a weighted Tchebycheff distance measure performed better than the Euclidean distance measure. The parameters were weighted inversely proportional to their variances. The results, however, were found to be relatively insensitive to the weighting coefficients. The additional hardware required for a typical microprocessor system, costs under $150. The ability to build the hardware for such a low cost was due to the use of Reticon's Universal Active Filter R5620, which costs under $7.00 each.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1983
- Accession Number
- ADA132475
Entities
People
- Walter F. Davidson
Organizations
- Air Force Institute of Technology