A Danish startup just showed that general-purpose AI gets roughly 1 in 6 words wrong on medical speech.
Copenhagen-based Corti published a research paper this week showing its clinical speech model, Symphony for Speech-to-Text, achieved a 1.4% word error rate on an English medical speech benchmark. OpenAI's Whisper large-v3 hit 17.7%. ElevenLabs came in at 18.1%.
The gap widens on the clinical details that actually change patient outcomes, things like medication dosages, measurements and dates. Symphony captured those correctly 98.3% of the time. The best general-purpose model managed 44.3%. That can be the difference between a correct instruction like "take 50mg twice daily" and an incorrect dose in the chart if the AI mishears it.
These benchmarks were designed by Corti, and no outside lab has independently verified the results. The company did open-source its evaluation tools and publish the full methodology though, so the claims are testable. Early real-world deployments are also consistent with these results.
At NHS Grampian's Donbank Ward in Scotland, nurses have been using Corti to document patient conversations in real time. Before the tool, staff would finish talking to a patient and try to reconstruct the conversation from memory. "Using the tech frees up mind space so we can have a really good conversation," Katie Anderson, a senior charge nurse at the ward, said. Her team even taught the system to handle their Scottish Doric accent, which she said "can confuse the system at times, but it's improving every day."
Staff reported that the technology improved the quality of patient interactions because they could actually be present in conversations instead of mentally cataloging everything for notes. They said the documentation got better, and that they felt this supported better patient care.
According to Corti's CTO Lars Maaløe, over 60 apps now focus on clinical scribing alone, and most of them, he says, rely on general-purpose AI. That may have seemed less critical when transcripts were used mainly as notes for a doctor's chart. But hospitals are now deploying AI agents for clinical decisions and health record navigation, and in many architectures, those systems reason from whatever the speech model gives them. If it gets a medication name wrong, the error can propagate into downstream systems unless caught by safeguards or human review.
Corti says it now serves over 100 million patients annually across health systems including the NHS, with platform sign-ups growing 30% quarter over quarter.
Corti's numbers are healthcare-specific, but the lesson isn't. If your business depends on accuracy in specialized language, whether legal, financial or clinical, your general-purpose AI may be producing error rates that go unnoticed without domain-specific benchmarks designed to catch them. Much of the AI industry conversation over the last three years has focused on which model scores highest on general benchmarks. Corti's work highlights that models which look "smart" on generic benchmarks can still make clinically dangerous errors on phrases like "metoprolol tartrate 50mg twice daily"—showing that in medicine, benchmark "smartness" and clinically relevant accuracy can diverge.
