University of Illinois Speech Accessibility Project

People with ALS, Parkinson's disease, cerebral palsy, Down syndrome, and stroke contribute paid voice recordings to the University of Illinois Speech Accessibility Project so that AI companies building voice interfaces can train systems that recognize their speech.

ENABLE Model location

• Builder-side Interventions → Create Accessible Content
• Builder-side Interventions → Set Requirements that Include Accessibility

What it is

The Speech Accessibility Project runs out of the Beckman Institute at the University of Illinois Urbana-Champaign, led by professor Mark Hasegawa-Johnson. The project recruits paid volunteers with speech differences to record voice samples using a web-based platform. Participants speak scripted prompts. UIUC researchers de-identify and curate those recordings into a shared dataset. Amazon, Apple, Google, Meta, and Microsoft fund the project and receive access to the dataset to train and improve their automatic speech recognition (ASR) systems.¹

As of June 2024, the project had collected approximately 415 hours of speech from more than 500 individuals with disabilities. By that date, 235,000 speech samples had been shared with the funding companies.² By September 2025, approximately 2,000 participants had contributed recordings.³

The project covers five diagnostic categories: Parkinson's disease, ALS, cerebral palsy, Down syndrome, and stroke. Community organizations including the Davis Phinney Foundation (Parkinson's community), Team Gleason (ALS community), and Lingraphica (augmentative and alternative communication) assist with participant recruitment and feedback.¹⁴

In January 2025, Microsoft reported that incorporating the project's data into training produced accuracy gains of 18% to 60% depending on the disability category. The improvements rolled out on Microsoft's Azure cloud endpoint for third-party customers.⁵

Why it matters

The corpora that shaped ASR development excluded atypical speakers for decades. The Switchboard corpus, assembled in the early 1990s from recorded telephone calls, skewed toward white speakers in the middle of the United States.⁶ DARPA-funded research through the 1990s and 2000s optimized for this population and embedded those conventions into benchmark standards. LibriSpeech, a large corpus assembled from public-domain audiobooks that became the dominant ASR benchmark through the 2010s, compounded the pattern. Its authors ranked speakers by word error rate and selected those with the lowest rates as the "clean" training split.⁷ The selection criteria structurally excluded atypical speech. When the Speech Accessibility Project launched, the largest existing database of atypical speech contained recordings from 16 people with cerebral palsy.⁵

The consequences of that default were measurable. Koenecke et al. tested five major ASR systems in 2020 (Amazon, Apple, Google, IBM, and Microsoft) against matched recordings from Black and white speakers. The average word error rate was 0.35 for Black speakers and 0.19 for white speakers. Among Black speakers, 23% of audio clips produced transcripts too degraded to use, compared to 1.6% for white speakers.⁶ For speakers with dysarthria, which affects 70-90% of people with Parkinson's disease and 80-95% of people with ALS at some point in their disease course, the gap was steeper.⁸⁹ ASR systems trained on audiobook narrators cannot learn the reduced loudness, imprecise articulation, shortened breath groups, and altered prosody that characterize dysarthric speech.

Social isolation, lost autonomy, and psychological distress are measurable costs of ASR failure for people with ALS and Parkinson's disease. When transcriptions fail repeatedly, disabled speakers have been documented blaming themselves for the breakdown rather than the technology. In a 2024 study of ALS patients using personalized ASR captioning, one participant attributed every transcription error to her own speech, leading researchers to warn of "potentially misplaced negative impacts on her own psychological well-being."¹⁰ Others withdraw from social contact rather than face the moment when the device fails in public. Before the SAP, when ASR failed, someone with ALS or Parkinson's disease had to repeat commands, hand over the device, or abandon the task. Communication failure in ALS is associated with social isolation, loss of control, changed social roles, and distress.⁹ These outcomes follow from a training data gap that no individual speaker can fix alone.

The SAP addresses that harm at the upstream source: builder-side content. By funding data collection from atypical speakers and making that data available to the companies that build ASR, the project shifts labor from the navigator to the builder. Disabled people record prompts once. The models improve for everyone who shares that speech pattern. This is a different structure than individual accommodation, where each person trains a personal voice model in isolation. The SAP creates shared infrastructure.

The institutional arrangement that produced the data gap also funds its correction, but unevenly. Amazon, Apple, Google, Meta, and Microsoft fund the SAP and receive priority access as founding members. As of April 2024, the project had also accepted 211 signed data use agreements from universities, nonprofits, and companies seeking access to the dataset.¹¹ That broader distribution differs from a proprietary model, but the data remains application-gated. Organizations that cannot navigate the agreement process, and open-source projects with no institutional affiliation, remain outside it. The five companies that built the training data conventions that excluded atypical speech are positioned to receive the most direct operational benefit: the accuracy gains roll out on their production systems, as Microsoft demonstrated on Azure. The adaptation tax falls not just on individuals but on any speech interface built outside the consortium. Smaller developers, public sector agencies, and community organizations that also serve disabled speakers inherit whatever gap the SAP does not yet reach.

The SAP represents the current frontier for systematic atypical speech data collection at industrial scale. Google's Project Euphonia, a parallel initiative, had collected 1.5 million utterances from approximately 3,000 speakers as of early 2025, achieving up to 85% improvement in word error rate in select domains compared to ASR models trained only on typical speech.¹² Together, these projects are establishing that atypical speech representation is an engineering requirement. What remains structurally unresolved is access: the data, the models, and the accuracy gains remain inside proprietary systems.

Real-world examples

In the news

Speech Accessibility Project data leads to recognition improvements on Microsoft Azure (January 2025)
-- EurekAlert!

• Microsoft reported 18% to 60% accuracy gains on atypical speech after training on Speech Accessibility Project recordings. The improvements applied specifically to speech affected by Parkinson's disease, ALS, cerebral palsy, Down syndrome, and stroke. The changes rolled out on Azure's cloud endpoint, meaning third-party developers building voice applications on Microsoft's platform inherit the gains. This is a builder-side change that propagates through the ecosystem without requiring individual disabled users to adapt.

In the news

University of Illinois joins five technology industry leaders in new Speech Accessibility Project (2022)
-- EurekAlert!

• The project launched with Amazon, Apple, Google, Meta, and Microsoft as funders. Before launch, the largest database of atypical speech included 16 people with cerebral palsy. The project addressed what Microsoft called a "data desert," the absence of diverse training data that left ASR systems calibrated for standard speech only. Participant contributions are paid, treating disabled speakers as data producers rather than subjects.

In the news

Speech Accessibility Project expands to Canada (October 2024)
-- Newswise / University of Illinois

• The SAP extended recruitment to Canadian participants with Parkinson's disease, ALS, cerebral palsy, Down syndrome, and stroke. The expansion reflects both the project's growth and the absence of comparable atypical speech data in Canadian English. Disabled speakers in Canada carrying the same ASR failure costs as American participants had no equivalent infrastructure for correction until the SAP extended its reach. The expansion does not address French-speaking Canadians, where the data gap continues.

In the news

Speech Accessibility Project now sharing recordings and data (April 2024)
-- Newswise / University of Illinois

• The SAP announced it had accepted 211 signed data use agreements and begun distributing Parkinson's speech data to universities, nonprofits, and companies. This shifts the project from a funding-members-only model toward broader research infrastructure. The distribution does not reach open-source projects or researchers without institutional backing, but it extends the dataset beyond the five founding companies for the first time.

• When the SAP launched, ASR systems trained predominantly on audiobook narrators. LibriSpeech, the dominant benchmark corpus, was built by selecting speakers with the lowest word error rates, structurally filtering out the speech variation that disability and accent introduce.⁷ The Switchboard corpus that preceded it, assembled from telephone calls in the early 1990s, skewed toward white speakers in the American Midwest.⁶

• Koenecke et al. (2020) found that all five major ASR systems produced a word error rate of 0.35 for Black speakers versus 0.19 for white speakers. Among Black speakers, 23% of transcripts were effectively unusable. The same structural cause, training data that underrepresented speaker diversity, produced both the racial and disability access gaps.⁶

• 70-90% of people with Parkinson's disease develop motor speech impairment as the disease progresses. Communication breakdown is associated with reduced participation, social withdrawal, and depression.⁸ Standard ASR cannot accommodate the acoustic changes that Parkinson's introduces: reduced volume, imprecise consonants, and variable prosody.

• 80-95% of ALS patients develop dysarthria at some point in their disease course. Researchers found that ASR accuracy for this population correlates with dysarthria severity, and that the functional threshold at which accuracy becomes usable remains poorly established.¹³⁹ In a 2024 study, actual word error rates for ALS speakers using personalized ASR captioning ran 38-53%, far above the system's own forecast of 15-35%, and meaning preservation rates ranged from 13-50%.¹⁰

• Google's Project Euphonia, which launched in 2018, ran parallel to the SAP. As of 2025 it had collected 1.5 million utterances from approximately 3,000 speakers and reported up to 85% improvement in word error rate on atypical speech in select domains.¹² Unlike the SAP, Project Euphonia stopped recruiting new participants and keeps its dataset within Google rather than distributing it via data use agreements.

What care sounds like (builder-side interventions)

Care at requirement-setting and content involves treating atypical speech as a design requirement from the start:

• "Our speech models must perform for the speakers who depend on voice interfaces most, including people with dysarthria, Parkinson's, and ALS."
• "We are funding data collection from people with ALS, Parkinson's, and cerebral palsy before we ship, not after."
• "We pay participants for their recordings because their speech is a technical resource and they are creating infrastructure."
• "We track word error rates across disability categories and publish them the way we publish benchmark scores."

What neglect sounds like (builder-side interventions)

• "We trained on LibriSpeech. It's the standard."
• "Our system handles most users fine."
• "We'll look at accessibility edge cases in a later release."
• "Users with severe speech impairments are a small market segment."
• "If the system doesn't understand them, they can use the keyboard."

What compensation sounds like (navigator-side compensations)

• "I don't talk to my neighbours much now I don't want that look of sympathy."¹⁰
• "When we walk as a group walk ahead so no one has to talk to me."¹⁰
• "This is my fault when I'm not understood by Relate."¹⁰
• "I have to repeat everything three times and even then it gets it wrong, so I just hand my phone to my wife."
• "I trained my own voice model but it only works on one device and now the software updated and I have to start over."
• "I stopped using voice assistants entirely. The error messages are humiliating and I don't have the energy for it anymore."
• "My speech changes as the day goes on. Mornings are better. So I do all my voice tasks in the morning before I'm too tired."
• "I type everything. It takes longer but at least it works."

---

All observations occur within the context of automatic speech recognition development and deployment in the United States, Canada, and Puerto Rico, as documented through the Speech Accessibility Project at the University of Illinois Urbana-Champaign.

Footnotes

1. University of Illinois joins five technology industry leaders in new Speech Accessibility Project -- EurekAlert!, 2022 ↩ ↩²

2. The Speech Accessibility Project: Best Practices for Collection and Curation of Disordered Speech -- Illinois Experts ↩

3. The Interspeech 2025 Speech Accessibility Project Challenge -- arXiv, 2025 ↩

4. University of Illinois Speech Accessibility Project -- Lingraphica ↩

5. Speech Accessibility Project data leads to recognition improvements on Microsoft Azure -- EurekAlert!, January 2025 ↩ ↩²

6. Racial disparities in automated speech recognition -- Koenecke et al., PNAS 2020 ↩ ↩² ↩³ ↩⁴

7. LibriSpeech: An ASR Corpus Based on Public Domain Audio Books -- Panayotov et al., IEEE ICASSP 2015 ↩ ↩²

8. Speech dysfunction, cognition, and Parkinson's disease -- PMC ↩ ↩²

9. Communication Matters -- Pitfalls and Promise of Hightech Communication Devices in Palliative Care of Severely Physically Disabled Patients With ALS -- PMC ↩ ↩² ↩³

10. How People Living With Amyotrophic Lateral Sclerosis Use Personalized Automatic Speech Recognition Technology to Support Communication -- Journal of Speech, Language, and Hearing Research, 2024 ↩ ↩² ↩³ ↩⁴ ↩⁵

11. Speech Accessibility Project now sharing recordings, data -- Newswise, April 2024 ↩

12. Project Euphonia: advancing inclusive speech recognition through expanded data collection and evaluation -- Frontiers in Language Sciences, 2025 ↩ ↩²

13. The use of speech recognition technology by people living with amyotrophic lateral sclerosis: a scoping review -- Disability and Rehabilitation: Assistive Technology, 2021 ↩