An Evaluation of Speech Recognition Technology.
Abstract
The last 15 years has seen the development of speech technology at very rapid rate. Unfortunately, the fact and fiction of recognition are not always easily separated. This confusion is not only evident among users, but also often among system integrators. This paper outlines the technology today and provides results from one set of benchmark tests. Three tests were performed with live speakers in three noise environments. The tasks used: a) a sixteen word discrete vocabulary, b) a 37 word connected speech vocabulary, and c) a 30 word connected speech vocabulary which was very tightly constrained (syntaxed). The noise environments included a 'quiet' background noise, noisy background of loud voices, and the sounds associated with a normal (loud) vehicle repair shop. The results indicated that virtually all systems tested could be made to perform well with specific, well-motivated speakers and under all noise conditions. Conditions not requiring advanced features (e.g., large vocabularies or connected speech) may turn these features into a liability through increased error. In spite of this, however, the technology is sufficiently mature to support many field applications. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 05, 1986
- Accession Number
- ADA179762
Entities
People
- Michael G. Joost
- Robert D. Rodman
- Taryn S. Moody
Organizations
- Battelle Memorial Institute