Speech Accessibility Project — Open Disordered-Speech Dataset Cuts Parkinson's Speech Recognition Errors

The Speech Accessibility Project (SAP) is led by researchers at the University of Illinois Urbana-Champaign's Beckman Institute in collaboration with five technology companies, including Google, and works with advocacy groups such as Team Gleason and the Davis Phinney Foundation. Its aim is to build a shared, consented dataset of impaired speech, since mainstream speech recognition is trained mostly on typical speakers and often fails people with conditions such as Parkinson's disease.
The first peer-reviewed results (Hasegawa-Johnson et al., Journal of Speech, Language, and Hearing Research, 2024) used the October 2023 data package: speech from 211 people with dysarthria related to Parkinson's disease, about 151 hours of recordings, with a test set of 42 additional speakers. A baseline recogniser trained on typical speech had a word error rate of 36.3%; after fine-tuning on SAP data it reached 23.7%, a drop of about 12.6 percentage points (roughly a third in relative terms). Error rates fell for almost all individual test speakers, with modest gains even for atypical patterns such as very fast speech or stuttering.
The recordings are available free to researchers, nonprofits and companies, so improvements are not locked inside one vendor. Google separately reports that its earlier Project Euphonia reduced median word error rates by more than 80% with personalised models, a company-reported figure not verified here.
Limits: the published result covers one condition (Parkinson's) and English, a few hundred speakers, a single published model setup, and word error rate on test data rather than measured change in users' daily lives. Other conditions are being collected but were not evaluated in the cited paper.

Read the full analysis: https://speechaccessibilityproject.beckman.illinois.edu/