Peer-reviewed research

Learning Curve of Speech Recognition

Tomi A. Kauppinen et al., Journal of Digital Imaging, 2013 · DOI 10.1007/s10278-013-9614-7

Read on October 16, 2026

What it is: A log-based study of radiologists at one Helsinki hospital who switched to speech recognition for writing reports. It compares characters corrected per report and unknown-word rates for 71 radiologists new to the system with 11 experienced users, and follows how the new users' error rates changed over time.
How much weight it carries: Peer-reviewed and based on system logs covering about 0.25 million reports, which is large. It is one hospital, one speech system and one clinical task, with 2000s recognition engines, so the numbers show the shape of the learning curve rather than what a new user will see with a current app.

Facts taken from this source

In the same 2013 study, experienced radiologists corrected 6 characters per report.
“Experienced radiologists corrected 6 characters for each report” (Abstract)
In a 2013 hospital study, new users of speech recognition corrected 11 characters per report, on average.
“Experienced radiologists corrected 6 characters for each report and for new users, 11.” (Abstract)
At the start, new users of speech recognition corrected 15 characters in a report of 433 characters.
“At the start, the new users corrected 15 characters (of 433)” (Results (character-based error rates))
Some new radiologists had a very unfavorable learning curve, with error rates that did not fall as expected.
“Some users presented a very unfavorable learning curve, with error rates not declining as expected.” (Abstract)
New users' rate of words unknown to the recognizer leveled off at 0.36 per report after about 10 weeks.
“with a plateau of 0.36 (38 % more than the experienced users) achieved after approximately 10 weeks” (Results (unknown words))

Peer-reviewed research

Tomi A. Kauppinen et al., Journal of Digital Imaging, 2013 · DOI 10.1007/s10278-013-9614-7

Read on October 16, 2026

What it is: A log-based study of radiologists at one Helsinki hospital who switched to speech recognition for writing reports. It compares characters corrected per report and unknown-word rates for 71 radiologists new to the system with 11 experienced users, and follows how the new users' error rates changed over time.
How much weight it carries: Peer-reviewed and based on system logs covering about 0.25 million reports, which is large. It is one hospital, one speech system and one clinical task, with 2000s recognition engines, so the numbers show the shape of the learning curve rather than what a new user will see with a current app.

In the same 2013 study, experienced radiologists corrected 6 characters per report.
“Experienced radiologists corrected 6 characters for each report” (Abstract)
In a 2013 hospital study, new users of speech recognition corrected 11 characters per report, on average.
“Experienced radiologists corrected 6 characters for each report and for new users, 11.” (Abstract)
At the start, new users of speech recognition corrected 15 characters in a report of 433 characters.
“At the start, the new users corrected 15 characters (of 433)” (Results (character-based error rates))
Some new radiologists had a very unfavorable learning curve, with error rates that did not fall as expected.
“Some users presented a very unfavorable learning curve, with error rates not declining as expected.” (Abstract)
New users' rate of words unknown to the recognizer leveled off at 0.36 per report after about 10 weeks.
“with a plateau of 0.36 (38 % more than the experienced users) achieved after approximately 10 weeks” (Results (unknown words))