Voice dictation and voice assistants make many more errors on speech affected by conditions such as Parkinson's disease, ALS, cerebral palsy, Down syndrome, and stroke, because the models were trained mostly on typical speech. The Speech Accessibility Project at the University of Illinois Urbana-Champaign is collecting recordings from adults with these conditions to close that gap, and its first results show large but uneven gains. For students whose speech is not recognized, the findings are promising for inclusion, yet the project's data comes from adults, so classroom performance remains untested.
Key takeaways
- At the end of 2023, speech recognition errors were under 2 percent for typical speech and over 18 percent for speech affected by dysarthria, according to the project's 2025 Interspeech paper.
- The project has recruited more than 1,400 contributors across five conditions, including 141 with Down syndrome and 316 with cerebral palsy.
- A Parkinson's-trained model cut word error from 36.3 percent to 23.69 percent on a test set, but speakers with atypical patterns gained only modestly.
- A public challenge lowered the best error rate on the project's data from 18 percent to 8 percent.
- No source reviewed here measured these tools with children in schools.
Why standard dictation fails some speakers
Automatic speech recognition learns from recorded examples. When the examples come from speakers whose voices are fluent and clear, the system learns those patterns well and treats other patterns as noise. Dysarthria, a motor speech disorder, can slow speech, slur consonants, or change volume and pitch. Down syndrome, cerebral palsy, stroke, ALS, and Parkinson's disease can all produce speech like this. The gap matters for inclusion because speech is often the easiest input for a person who also has a physical impairment. The project's authors make this point directly, noting that people with dysarthric speech often have physical impairments that make accurate recognition more valuable for daily activities.
The measure used is word error rate: the share of words that are wrong, missing, or added when the system's transcript is compared with a human transcript. According to the project's 2025 Interspeech paper, the rate at the end of 2023 was below 2 percent for non-dysarthric speech and above 18 percent for dysarthric speech.
How the Speech Accessibility Project collects data
The Speech Accessibility Project is based at the university's Beckman Institute. Its funders include Amazon, Apple, Google, Meta, and Microsoft, plus the U.S. National Institute on Deafness and Other Communication Disorders and the National Science Foundation. Participants are adults. Mentors, typically licensed speech-language pathologists, screen each person by video call to confirm that their speech meets the criteria for one of the five conditions, and participants then record at home on a computer or phone, reading prompts, giving voice commands, and answering open questions. A team of more than 20 annotators transcribes every recording.
The project set goals of 1,000 hours of speech and 2,000 participants, with 400 for each condition. Its 2025 paper reported the following contributor counts with six months of collection remaining, and said the 1,000-hour goal was passed in early 2025.
| Condition | Contributors reported | Share of the 400 goal |
|---|---|---|
| Parkinson's disease | 442 | Above goal |
| ALS | 334 | About 84 percent |
| Cerebral palsy | 316 | 79 percent |
| Stroke | 190 | About 48 percent |
| Down syndrome | 141 | About 35 percent |
The last column is a calculation from the reported counts and the stated goal. Down syndrome had the lowest count, which matters for readers thinking about students, since the conditions with the most data are likely to see the largest gains first.
What the first results show
In a study reported by the Beckman Institute on September 27, 2024, and published in the Journal of Speech, Language, and Hearing Research, a recognizer trained on Parkinson's-related speech reached a word error rate of 23.69 percent on a new test set, against 36.3 percent for a comparison model trained on speech from people without the condition. Training used 151 hours from 190 participants. The authors called the findings initial, and error rates fell for almost all test speakers, while those with unusually fast speech or stuttering gained only modestly.
A later public challenge built on the data. The project's challenge page reports that the best error rate fell from 18 percent, for the best open-source system before the competition, to 8 percent for the winner's best score in February 2025. Because the first study used Parkinson's speech only, its figures do not transfer automatically to other conditions.
What is not yet known for schools
Several gaps separate these results from a classroom decision.
- Age: the project recruits adults, and children's voices differ in pitch and articulation, so a model's accuracy for students is unmeasured here.
- Conditions: Down syndrome and cerebral palsy have fewer recordings than Parkinson's disease, and the 2024 study reported only on Parkinson's.
- Setting: recordings were made at home with prompts, not in a noisy room with several speakers.
- Access: data is shared with funding companies and with applicants approved by the project, so a given school product may or may not have used it.
- Support: accuracy figures say nothing about whether a tool improves learning or participation.
Whether a tool suits a particular student is a matter for that student, the family, and the school team, ideally with a speech-language pathologist. This article gives general information and is not a clinical or educational recommendation.
Questions to ask a vendor
- Was the recognizer tested on speakers with the relevant condition, and on children?
- What word error rate was reported for those speakers, compared with typical speakers?
- Can the tool be adapted to one person's voice, and where are the recordings stored?
- What happens when it fails, and is there a typed or switch-based fallback?
FAQ
Who funds the Speech Accessibility Project?
Amazon, Apple, Google, Meta, and Microsoft, along with a U.S. National Institutes of Health institute and the National Science Foundation, according to the Beckman Institute's 2024 release.
Does the project record children?
Its published description concerns adults. Whether children's speech will be added was not found in the sources reviewed.
Is word error rate the same as understanding?
No. It counts word-level mistakes against a human transcript. A transcript can score well and still miss meaning, tone, or intent.
Can a school use the data directly?
The data is available through an application process for researchers, nonprofits, and companies. Schools are more likely to meet it indirectly through products built by those groups.
Related reading
- AI in special education over a decade
- The impact of AI technology in education
- Enhancing learning with AI
- AI chatbots in schools
More in AI and Society.
Last verified: October 2026
Sources
- Speech Accessibility Project, Beckman Institute, University of Illinois Urbana-Champaign, 2026
- Zwilling et al., The Speech Accessibility Project: Best Practices for Collection and Curation of Disordered Speech, Interspeech 2025
- Automatic speech recognition learned to understand people with Parkinson's disease by listening to them, Beckman Institute, September 27, 2024
- Community-Supported Shared Infrastructure in Support of Speech Accessibility, Journal of Speech, Language, and Hearing Research, 2024
- Speech Accessibility Project challenge competition, Beckman Institute, 2026