This is the repo of "Decoding Disparities: Evaluating ASR System Performance in Transcribing Black and White Patient Verbal Communication with Nurses in Home Healthcare". This study aims to analyze the performance of ASR systems verbal communication of Black and White patient-nurse encounters in home healthcare as the foundation for enhancing integration of this data into EHR and influencing patient care. The first aim is to evaluate the accuracy of four commercial and open-source ASR systems in transcribing audio-recorded patient-nurse verbal communication in home healthcare settings. The second aim is to investigate potential differences in the accuracy of these systems when transcribing verbal communications between older adult patients and nurses, with a particular focus on differences between Black and White patients.
The dataset comprises audio recordings of patient-nurse communications gathered in the course of evaluations taking place in a large, established home healthcare service based in New York City. Conversations were on average 15 minutes in duration (range 11 to 21 minutes) with a median of 1,186 words spoken per patient. Overall, 860 utterances were available for study, including 475 drawn from Black patients and 385 from White patients.
While AWS systems showed superior accuracy in home healthcare, all ASR systems revealed substantial differences by race, underscoring the urgent need for more diverse training datasets and improved dialect sensitivity in ASR technology
For questions or inquiries, please contact m.zolnoori@gmail.com