Healthcare software is moving toward a future where clinicians can simply say what they need. Document this encounter. Show me the latest CBC. Pull up the relevant imaging. Start an order. Instead of navigating a maze of screens, clinicians can interact with clinical software and data through natural language.

There is good reason to want that future. Physicians spend an average of 13 hours each week on indirect patient care, including documentation, order entry, test results, and referrals. Conversational technology offers an opportunity to take some of that work off their plates.

But before a system can reliably act on what a clinician means, it has to capture what they said and understand the context surrounding those words. Only then can we expect it to translate speech into intent and action. Healthcare cannot skip those steps.

Speech recognition has been part of healthcare for decades, but clinical speech remains particularly difficult to get right. Medical terminology, drug names, dosages, measurements, specialty-specific language, different accents, and clinical environments all create opportunities for error.

A 2025 systematic review of 29 studies of AI-based speech recognition for clinical documentation found word error rates ranging from 8.7% in controlled dictation settings to more than 50% in conversational or multi-speaker environments. A 2026 study similarly found higher error rates for non-native English speakers across three speech recognition approaches, including an 11 percentage point increase in word error rate with Whisper.

Word error rate only tells us so much, though. A transcript can be nearly perfect and still get the one thing that matters wrong.

Source link