Peer-reviewed research

The use of generative artificial intelligence-based dictation in a neurosurgical practice: a pilot study

Benjamin S. Hopkins et al., Neurosurgical Focus, 2025 · DOI 10.3171/2025.4.FOCUS24834

Read the original

Read on October 16, 2026

What it is: A pilot study in which three physicians dictated neurosurgical procedures, processed by a modified Whisper-based model and by Nuance's Dragon Medical One. It compares word error rates for the two systems, with and without linguistic errors.
How much weight it carries: Peer-reviewed journal article, but a pilot with three dictating physicians and one specialty, so it cannot show general accuracy. The AI model was modified and is not stock Whisper. Authors are listed as the first three of twelve.

Facts taken from this source

The study used Nuance's Dragon Medical One as its commercial comparison.
“Nuance's Dragon Medical One as a comparative commercial standard” (Methods)
The study's AI system was a modified speech-to-text model built on the backbone architecture of OpenAI's Whisper.
“a modified speech-to-text model implemented based on a backbone architecture created by OpenAI's Whisper model” (Methods)
The pilot study had three independent physicians dictate and record the neurosurgical procedures.
“neurosurgical procedures were dictated and recorded by three independent physicians” (Methods)
In the pilot, the difference in mean word error rate across all errors was not statistically significant.
“The mean WER was 1.75% for Whisper and 1.54% for Dragon (p = 0.080).” (Results)
In the same pilot, Dragon Medical One had a mean word error rate of 1.54 percent, counting all errors.
“The mean WER was 1.75% for Whisper and 1.54% for Dragon (p = 0.080).” (Results)
When linguistic errors were excluded, Dragon Medical One had a mean word error rate of 1.34 percent.
“When excluding linguistic errors, Whisper outperformed Dragon with a mean WER of 0.50% versus 1.34%” (Results)
In the 2025 neurosurgery pilot, the modified Whisper-based model had a mean word error rate of 1.75 percent, counting all errors.
“The mean WER was 1.75% for Whisper and 1.54% for Dragon (p = 0.080).” (Results)
When linguistic errors were excluded, the Whisper-based model had a mean word error rate of 0.50 percent.
“When excluding linguistic errors, Whisper outperformed Dragon with a mean WER of 0.50% versus 1.34%” (Results)

The use of generative artificial intelligence-based dictation in a neurosurgical practice: a pilot study

Facts taken from this source

Pages that rest on this source

The use of generative artificial intelligence-based dictation in a neurosurgical practice: a pilot study

Facts taken from this source

Pages that rest on this source