Fact · Accessibility and health
4.6percent word error rate
In Google's Project Euphonia study of 432 people with disordered speech, personalized speech recognition models reached a median word error rate of 4.6 percent, against 31 percent for general speaker-independent models.
Checked against its source on October 16, 2026
In the source's own words
“The accuracy of personalized models for recognizing disordered speech was high (WER: 4.6%), and significantly better than speaker-independent models (WER: 31%).”
Automatic Speech Recognition of Disordered Speech: Personalized models outperforming human listeners on short phrases, Jordan R. Green et al., Interspeech 2021 (Google Research publication page), 2021, Abstract, Results.
Open the original- What it measures
- Each participant recorded at least 300 short phrases through a web app, and a model was trained on each person's own voice. The personalized models also outperformed expert human transcribers, with the biggest gains for the most severely affected speakers.
- What it does not show
- Short phrases, not open dictation, and each person had to record hundreds of phrases first. Low signal-to-noise ratio and fewer training phrases made results worse.
Source
Cited on
- Speech to Text for Accessibility, RSI and DisabilitiesTopic
- How Speech Recognition Works, From Sound to TextGuide
- How Accurate Is Speech Recognition? The Real NumbersGuide
- How do you improve Mac dictation accuracy?Question
- Is speech recognition less accurate for some people?Question
- What is speech recognition used for?Question