The Speech Accessibility Project is Teaching AI to Understand Impaired Speech

Voice assistants, dictation software, and voice activated devices have become everyday tools, but they only work if the system can understand you. For millions of people whose speech differs from the norm because of Parkinson's disease, amyotrophic lateral sclerosis, stroke, cerebral palsy, or Down syndrome, these systems simply fail. The Speech Accessibility Project, led by the University of Illinois Urbana-Champaign, is fixing that by building the datasets AI companies need to understand everyone.

The project, which launched in 2022 with backing from Amazon, Apple, Google, Meta, and Microsoft, pays people with impaired speech to record phrases and sentences. Those recordings form a private, de-identified dataset that participating companies use to retrain their speech recognition models. As of the end of September 2025, about 2,000 participants had submitted recordings, and the current data package includes 1,500 hours of speech from roughly a thousand participants. The results are already measurable. Microsoft has reported accuracy gains between 18 and 60 percent on impaired speech after training with project data.

Why Standard Speech Recognition Fails

Speech recognition systems learn from enormous datasets of recorded speech. If the training data contains mostly clear, typical voices, the resulting model will struggle with anything else. Accents have long been a known weakness. Impaired speech, sometimes called dysarthric speech, is a far bigger gap. Word error rates for dysarthric speech remained stubbornly high even as recognition of typical speech approached human parity.

The consequences go beyond convenience. Voice control is often the most accessible way to interact with technology for people with limited mobility. When a phone, smart speaker, or dictation tool cannot understand the user, it can shut them out of communication, independence, and work. The problem is not that the AI cannot be trained. The problem is that nobody had collected enough impaired speech data to train it well.

Paying Participants to Build the Dataset

The Speech Accessibility Project recruits adults in the United States and Canada with Parkinson's disease, ALS, Down syndrome, cerebral palsy, and speech effects of stroke. Participants record prompted phrases using a simple web interface from home, and they are paid for their time. This design matters. Collecting impaired speech is slow and effortful for the speaker, and compensating participants respects that effort while producing a dataset with genuine diversity of conditions, severity levels, and demographics.

By June 2024 the project had shared 235,000 speech samples with the companies funding it. The research community received its first partial dataset release in April 2024, containing more than 400 hours of speech from over 500 individuals with disabilities. That release powered the Interspeech 2025 Speech Accessibility Project Challenge, a public competition in which research teams from around the world competed to build the most accurate recognition systems for impaired speech.

Measurable Gains in the Real World

The corporate impact is concrete. Microsoft reported accuracy improvements between 18 and 60 percent after incorporating project data into its speech models. Google's related efforts grew out of Project Euphonia, a research initiative that collected over a million speech samples from people with impaired speech and led to Project Relate, an Android app that transcribes impaired speech in real time and restates it in a clear synthesized voice.

The project has also opened its data to nonprofits and companies beyond the original founders through a proposal process. Any organization building speech recognition tools can apply to use the dataset, which means the benefits can spread to augmentative communication devices, accessibility software, and niche products that serve small populations the biggest companies would never prioritize on their own.

A Model for Inclusive AI

What makes this project unusual is its structure. Competing companies, including Amazon, Apple, Google, Meta, and Microsoft, fund a shared dataset that any of them can use, alongside academic researchers. The participants who contribute the data are paid rather than mined. The dataset is designed to be private and de-identified from the start. It is a working template for how the AI industry can close representation gaps it cannot solve alone.

There is a lesson here that reaches beyond speech. Every AI system reflects its training data, and groups missing from that data get worse results. Fixing that does not require a breakthrough. It requires the unglamorous work of recruiting participants, recording sentences, compensating people fairly, and sharing the results. A thousand people recording phrases from their living rooms have already made voice technology measurably better for millions.