Technology Evaluation: Comparative Evaluation of Three Continuous Speech Recognition Software Packages in the Generation of Medical Reports

OBJECTIVE To compare out-of-box performance of three commercially available continuous speech recognition software packages: IBM ViaVoice 98 with General Medicine Vocabulary; Dragon Systems NaturallySpeaking Medical Suite, version 3.0; and L&H Voice Xpress for Medicine, General Medicine Edition, version 1.2. DESIGN Twelve physicians completed minimal training with each software package and then dictated a medical progress note and discharge summary drawn from actual records. MEASUREMENTS Errors in recognition of medical vocabulary, medical abbreviations, and general English vocabulary were compared across packages using a rigorous, standardized approach to scoring. RESULTS The IBM software was found to have the lowest mean error rate for vocabulary recognition (7.0 to 9.1 percent) followed by the L&H software (13.4 to 15.1 percent) and then Dragon software (14.1 to 15.2 percent). The IBM software was found to perform better than both the Dragon and the L&H software in the recognition of general English vocabulary and medical abbreviations. CONCLUSION This study is one of a few attempts at a robust evaluation of the performance of continuous speech recognition software. Results of this study suggest that with minimal training, the IBM software outperforms the other products in the domain of general medicine; however, results may vary with domain. Additional training is likely to improve the out-of-box performance of all three products. Although the IBM software was found to have the lowest overall error rate, successive generations of speech recognition software are likely to surpass the accuracy rates found in this investigation.