Fact · Accessibility and health
23.7percent word error rate (from 36.3)
A baseline speech recognizer with a 3.4 percent word error rate on typical speakers had a 36.3 percent word error rate on speech from people with Parkinson's disease; fine-tuning on Speech Accessibility Project data reduced it to 23.7 percent.
Checked against its source on October 16, 2026
In the source's own words
“A baseline ASR, with a word error rate of 3.4% for typical speakers, transcribes test speech with a word error rate of 36.3%. Fine-tuning reduces the word error rate to 23.7%.”
Community-Supported Shared Infrastructure in Support of Speech Accessibility, Mark Hasegawa-Johnson et al., Journal of Speech, Language, and Hearing Research, 2024, Abstract, Results.
Open the original (DOI 10.1044/2024_JSLHR-24-00122)- What it measures
- The test used the project's 2023-10-05 data package, with speech from 211 people with dysarthria associated with Parkinson's disease for training and 42 other speakers for testing. The gap between 3.4 and 36.3 percent shows how much worse a typical recognizer does on dysarthric speech.
- What it does not show
- An open-source research model (wav2vec 2.0 based), not a commercial dictation product, and a first data release. Even after fine-tuning, roughly 1 word in 4 was wrong.
Source
Cited on
- Speech to Text for Accessibility, RSI and DisabilitiesTopic
- Word Error Rate (WER): How to Calculate and Read ItGuide
- How Accurate Is Speech Recognition? The Real NumbersGuide
- Speech Recognition for Stuttering and DysarthriaGuide
- Can voice recognition understand dysarthric speech?Question
- Is speech recognition less accurate for some people?Question
- What is a good word error rate?Question
- DysarthriaGlossary