Voice for All: Evaluating the Accuracy and Equity of Automatic Speech Recognition Systems in Transcribing Patient...

Integrating automatic speech recognition (ASR) systems into home healthcare workflows can enhance risk prediction models. However, ASR systems exhibit disparities in transcription accuracy across racial and linguistic groups, highlighting an equity gap that could bias healthcare delivery. We evaluated four ASR systems-AWS General, AWS Medical, Whisper, and Wave2Vec-in transcribing 860 patient-nurse utterances (475 Black, 385 White). Word error rate (WER) was the primary measure. AWS General achieved the highest accuracy (median WER 39%), but all systems were less accurate for Black patients, particularly in linguistic domains "Affect," "Social," and "Drives." AWS Medical outperformed others on medical terms, although filler words, repetition, and nonmedical terms challenged every system. These findings underscore the need for diverse training datasets and improved dialect sensitivity to ensure equitable ASR performance and robust risk identification in home healthcare.