Recognition accuracy with a voice-recognition system designed for anesthesia record keeping

We tested on three occasions, with anesthetists as subjects, the accuracy of two voice-recognition systems designed for anesthetic record keeping. Initially, a prototype system was tested (10 subjects); several years later the resulting commercial system was tested in a quiet environment (11 subjects) and in noisy operating rooms (10 subjects). For each test an anesthetist first trained the system to recognize his or her voice by reading aloud a list of common anesthetic terms. To determine recognition accuracy, the percentage of words recognized correctly by the computer, each subject repeated the vocabulary words ten times. Although accuracy was similar during the three tests, it was slightly higher with the laboratory test (mean percent of words recognized correctly, 96.5%; range of accuracy for individual anesthetists, 91.6 to 98.8%) than with the prototype test (95.9%; range, 89.1 to 99.6%). Accuracy was lowest with the operating room test (95.3%; range, 87.8 to 98.4%). Twenty-four words caused particular difficulty during the laboratory test and were eliminated from the vocabulary of the subsequent operating room test. Omitting these 24 words from the laboratory vocabulary list allowed a more nearly direct comparison with the results from the operating room list; recognition accuracy improved in the former to 97.5% (range, 92.1 to 98.9%). Two anesthetists--one each from the laboratory and operating room tests--performed poorly, and eliminating their scores changed the respective overall scores to 98.2% (range, 96.7 to 98.9%) and 96.5% (range, 94.3 to 98.4%). Thus, the corrected difference between the laboratory accuracy and the operating room accuracy was 1.7%.