Speech recognition at two field sites
Ann M. Rollins, B. Constantine, S. Baker · 1983
The performance of two speech recognition systems installed at two field sites was analyzed. The speech systems were part of larger computer systems that were performing real functions in industrial environments. The two sites appeared to be polarized in terms of expected suitability for speech recognition. The variables looked at included task complexity, memory load, requirements for verification and error correction, vocabulary and syntax, microphone, operator experience and complexity of host computer software. Accuracy and throughput were measured for the speech recognition system at each site. The same measurements were made for keyboard entry. Operator differences account for most of the variance in results. Accuracy with voice input was higher than with keyboard for most operators. The most accurate operators with keyboard also tended to be the most accurate with voice. Throughput data appears more sensitive to individual differences in dealing with voice input, although the throughput data was clouded by slow host system response times overall. The discussion suggests that one to one replacement of keyboard with voice overlooks some possible advantages of voice. It is also possible to find operators who work well with voice. For those who do not work well with voice, the problems appear to be related to general work habits and attitude, rather than to specific difficulties with speech.