Preprint
Using Confidence Scores to Improve Eyes-free Detection of Speech Recognition Errors
Sadia Nowrin and Keith Vertanen, arXiv, 2024
Read the originalRead on October 16, 2026
- What it is
- A study of how people who cannot see the screen detect speech recognition errors when the transcript is read aloud. It tests whether slowing the audio only where the recognizer was unsure helps listeners catch the errors.
- How much weight it carries
- Preprint (version 3, revised June 2025), not yet confirmed as a peer-reviewed publication in our search. The method is clearly described, but the study is a single lab experiment on read-back audio, so it shows what helps listeners, not what happens in everyday use.
Facts taken from this source
- In the Nowrin and Vertanen study, slowing the read-back audio only where the recognizer was unsure led to a 12 percent relative increase in how well listeners detected errors, compared with uniformly slowing the audio.
“selectively slowing down the audio when the recognizer exhibited uncertainty led to a 12% relative increase in participants' ability to detect errors compared to uniformly slowing the audio” (Abstract)
- Nowrin and Vertanen state that visual feedback may not always be practical for people who are blind or low-vision when they check recognized speech.
“While visual feedback can help detect errors, it may not always be practical, especially for people who are blind or low-vision.” (Abstract)