Medication Names Are a Safety-Critical Failure Mode in Clinical Speech Recognition: An Open Benchmark of On-Device and Cloud Systems

Doctors spend more than half their workday writing notes about their patients. This heavy administrative burden is a leading cause of burnout, pushing many physicians toward the edge of exhaustion. In recent years, a new technology has emerged to help: ambient artificial intelligence. These systems listen to conversations between doctors and patients and automatically turn the spoken words into written text, hoping to free up the physician's time. However, for this technology to be safe, the computer must understand exactly what was said. If the software mishears a medication name, the consequences could be severe. The core challenge lies in how these systems are built. Most current services send the audio recording over the internet to a massive server farm in the cloud to be processed. This raises concerns about patient privacy and the cost of sending that data back and forth. An alternative approach is to run the software directly on the doctor's device, like a laptop or phone, keeping the audio local and never sending it out. But until now, no one had a clear, open way to compare how well these local devices perform against the powerful cloud giants, especially when it comes to the most critical words in a medical conversation.

A researcher at Emory University School of Medicine set out to build that missing comparison tool. They created a standardized test using 400 simulated conversations between doctors and patients, covering three different medical specialties: respiratory issues, primary care, and psychiatric evaluations. These conversations were not real patient visits but were carefully recorded scenarios designed to mimic the flow of a real exam. The researcher then ran these same recordings through twelve different speech recognition models that can run on a local device, and five different commercial cloud services that process audio over the internet. The goal was to see which system could accurately capture the specific medical terms that matter most, such as drug names, symptoms, and procedures, without making dangerous mistakes.

The results revealed a clear and worrying pattern. While the local devices performed surprisingly well on general speech, they struggled significantly with medication names. In the simulated conversations, the local models misheard specific drug names in a high percentage of cases, ranging from 85% to 93% depending on the medication and dataset. For example, when the word "lisinopril" was spoken in the primary care dataset, the local software got it wrong in 92 percent of the cases, and "amoxicillin" in 93 percent. The cloud services did much better, getting those same names wrong only 3 to 30 percent of the time. This gap suggests that while the local computers are good at understanding general conversation, they lack the specific training to recognize the tricky, sound-alike names of medicines that are critical to patient safety. The study found that even the best local models were not yet reliable enough to be trusted without a human double-checking the drug names before the note is finalized.

The researcher also tested whether combining the outputs of multiple local models could fix this problem. The idea was that if one model missed a word, another might catch it, and a voting system could choose the correct answer. However, this approach offered very little help. Merging the results of different models improved the accuracy by less than one percentage point, a difference so small it would not change the outcome in a real-world scenario. The study showed that simply adding more models does not solve the core issue; the fundamental limitation lies in the models themselves, not in how they are combined. The researcher also noted that one specific model, designed for medical dictation, did perform much better on drug names than the others, suggesting that the problem can be solved if the software is specifically trained for medical use rather than general conversation.

Beyond accuracy, the study looked at the cost and privacy implications. Sending audio to the cloud costs money, ranging from a few cents to over a dollar per patient visit, depending on the service. For a doctor seeing thousands of patients a year, this adds up to thousands of dollars. In contrast, running the software on a local device costs nothing per visit after the initial setup, as it uses the computer the doctor already owns. This makes local processing a much more affordable option for smaller clinics or those in areas with limited internet access. Furthermore, keeping the audio on the device means the patient's private health information never leaves the room, addressing significant privacy concerns that arise when data is sent to third-party servers.

The study concludes that while local speech recognition is a viable and cost-effective alternative to cloud services, it is not yet perfect. The technology is close enough to be useful, but it cannot be trusted to handle medication names on its own. The researchers emphasize that any system using this technology, whether it runs on a local device or in the cloud, must include a mandatory step where a human reviews and confirms the drug names before the medical record is saved. Until the software improves specifically on these critical terms, the safety of the patient depends on that final human check. The work provides a clear roadmap for the future: rather than trying to combine many imperfect models, the focus should shift to training single models specifically on medical language to close the remaining gap in accuracy.