Speech recognition devices such as Google Home, Amazon Echo and Alexa, Siri, and Cortana assist a wide range of users with tasks from answering questions to managing smart homes. However, these tools are often inaccessible to individuals with speech differences or disabilities.
Automatic speech recognition (ASR) systems are trained using large datasets from public audiobook websites, resulting in technology that excels at transcribing voices similar to professional narrators but struggles with challenges like second-language accents, underrepresented dialects, and speech impacted by disabilities.
The Speech Accessibility Project (SAP) at the University of Illinois Urbana-Champaign’s (UIUC) Beckman Institute for Advanced Science and Technology aims to address this gap by collecting 1,000 hours of transcribed audio from people with speech-affecting disabilities.
“The goal is to make it possible for researchers to train speech recognizers that can cope with that wide range of atypical speech in order to correctly recognize what the person is saying,” says Mark Hasegawa-Johnson, PhD, project lead and professor of electrical and computer engineering at UIUC.
Identifying the importance of a standardized dataset for improving ASR systems, last year a consortium of companies including Amazon, Apple, Google, Meta, and Microsoft sought proposals from universities on a project to ethically collect samples from individuals with relevant disabilities. UIUC was chosen to lead this initiative, resulting in the formation of SAP.
As part of its current phase, SAP is actively enlisting individuals who have Parkinson’s disease to record speech samples. Participants are tasked with three types of prompts, including commands designed to study device interaction techniques, phonetically diverse sentences from novels to capture various speech sounds in a range of contexts, and spontaneous speech where they share personal opinions and preferences.
The project has garnered substantial interest, with over 750 individuals reaching out to volunteer. A speech pathologist evaluates each potential participant to assess their suitability for speech recognition studies, says Hasegawa-Johnson.
To date, just over 200 people have successfully completed the recording sessions. Participants are compensated and receive assistance with logging in and completing the process.
“We want to value the time of our volunteers as well as their caregivers,” says Clarion Mendes, clinical assistant professor at UIUC, speech-language pathologist, and SAP team member. Furthermore, to safeguard participant privacy, Mendes says the project shares only aggregated demographic data and requires organizations to submit an application before accessing the information.
SAP continues to actively seek volunteers with Parkinson’s disease, including college students, and plans to recruit individuals with ALS, stroke-related disabilities, Down syndrome, and cerebral palsy in upcoming phases.●
To learn more and sign up to volunteer, visit speechaccessibilityproject.beckman.illinois.edu.
This article was published in our November/December 2023 issue.