Effect of ambient sound masking on the accuracy of computerized speech recognition

Purpose: To evaluate the direct effect that sound masking has on computerized speech recognition (SR) accuracy in the digital reading room while eliminating the Lombard effect. Materials and Methods: This study complies with HIPAA requirements, and the institutional review board determined that approval was not required; informed consent was obtained. Ten radiologists digitally recorded 20 selected reports. Recorded reports were transcribed by using a commercial SR system at sound-masking levels 1-4, representing lowest to highest A-weighted sound pressure measurements in decibels. Dictated reports were compared with original reports to determine numbers of errors. A repeated-measures analysis of variance was used for overall comparison of mean percentage of transcription inaccuracies across sound-masking levels, and paired t tests were used to compare each sound-masking level to the baseline. Results: Mean percentage of transcription inaccuracies at baseline was 11.6% (range, 6.3%-26.1%). Mean percentages of transcription inaccuracies at sound-masking levels 1-4 were 11.3%, 12.3%, 13.0%, and 13.6%, respectively. Mean percentages of transcription inaccuracies differed across all sound-masking levels (P < .001) but not between baseline and sound-masking level 1 (P = .313). Mean percentages of transcription inaccuracies at sound-masking levels 2-4 were each significantly higher than at baseline (P < .01 for each). Conclusion: Low-level sound masking provided slightly but not significantly improved SR accuracy, whereas higher levels decreased transcription accuracy. Appropriate levels of sound masking in reading room environments may decrease the negative effect of ambient noise without a deleterious effect on SR accuracy.