ISSUES IN THE EVALUATION OF SPEECH RECOGNITION SYSTEMS

Authors
  1. Taylor, M.M.
Corporate Authors
Defence and Civil Inst of Environmental Medicine, Downsview ONT (CAN)
Abstract
Many factors come into play when Automatic speech recognition (ASR) devices are to be evaluated. Different kinds of information are required by designers, by manufacturers, and by potential purchasers. Some applications may require very accurate recognition of a few selected and trained talkers, whereas others require reasonably accurate recognition of untrained self-selected talkers. Error percentages are dependent on the test vocabulary, and may not even be monotonic with overall recognizer capacity. A measure called "Effective Vocabulary Capability" EVC is introduced. The EVC is conceptually independent of the vocabulary used to test the machine. The EVC of a device is the maximum size of vocabularly that the device could recognize with a tolerably small error rate, if the words in the vocabulary were chosen for maximum distinctiveness.
Report Number
DCIEM-86-P-16 — Research Paper
Date of publication
15 Jun 1986
Number of Pages
30
Reprinted from
J of the American Voice I/O Society, vol 3, 1986, p 39-67
DSTKIM No
87-01126
CANDIS No
49969
Format(s):
Hardcopy;Originator's fiche received by DSIS

Permanent link

Document 1 of 1

Date modified: