What causes speech recognizers to make mistakes?

Informed selections among the available commercial speech recognizers requires knowledge of their accuracies under representative test conditions, and evidence concerning what causes errors. An organized list of over 80 factors that might affect recognition accuracy includes fixed specifications of devices (e.g., type of device, number of frequency filters, bits/word in reference patterns) and controllable variables (e.g., speaking rate, microphone, number of speakers, speaker skill, sex, dialect, physical and pyschological state, vocabulary, form of feedback for verification, etc.). Very little evidence has been accumulated regarding the relative impact of these factors related to task constraints, human factors, languages for interaction, channel effects, algorithmic factors, types of errors, or response processes. Controlled tests have been defined for comprehensive assessment of the accuracies of isolated word recognizers operating with various speakers, training methods, vocabularies, database recording procedures, rates of data entry, noise and channel conditions, etc. A statistical pre-test and method of "response surface methodology" will simplify the tests.