Speech Accessibility Project

Overview

The goal of the Speech Accessibility Project at the University of Illinois Beckman Institute (https://speechaccessibilityproject.beckman.illinois.edu) is to collect, annotate, and curate a shared database of speech samples from people with atypical speech, and share this data set with researchers at other organizations. This two-year project plans to collect 1,200,000 speech samples from 2,000 people, each of whom will provide 600 samples. In Year 1, the initial focus will be people with Parkinson's. In Year 2, four more etiologies of interest will be recruited: Amyotrophic Lateral Sclerosis (ALS), Cerebral Palsy (CP), Down Syndrome (DS), and Stroke. UIUC will build an open-source software infrastructure to collect annotated speech samples and share these data in an appropriately secure fashion with researchers from our partner technology companies (and eventually, other organizations as well) so that they can use these data to improve their automatic speech recognition algorithms. This project promotes diversity, equity, and inclusion by helping technology companies to fully support all types of speech, and it is also more efficient and less burdensome for these specialized patient populations to have one centralized "collector" of speech samples.

The goal of our project is to collect 1,200,000 speech samples from 2,000 people with dysarthria, where we expect to collect data from 400 people each from five different patient populations. Each person would provide 600 speech samples. (600 samples/person x 400 persons/etiology x 5 etiologies = 1,200,000 samples) Our schedule of research procedures is: 1. February or March-August 2023: data collection of speech samples from 400 people with Parkinson's. 2. August 2023-August 2025: data collection of speech samples from 1,600 people with ALS, CP, DS, and Stroke. Data collection of speech samples in Year 1 will be a collaboration of the University of Illinois and of mentors from Lee Silverman Voice Therapy (LSVT) Global. Potential participants will be screened both with a questionnaire and by providing a short set of "quality control" speech samples. If the participant does not pass screening, they will be thanked for their interest. Otherwise, the participant is eligible for the study and can do the informed consent process and then engage in contributing speech samples. Participants can do as many recordings as they wish at whatever time of day is convenient for them. Participants will be able to login to the system at any time, 24/7. In Year 2, this procedure will be performed with patients from other etiologies with additional advocacy organizations as partners. Participants who are unable to read text from the computer screen will be offered the opportunity to record speech using a verbal-repetition protocol. In order to participate in the verbal repetition protocol, a participant must be accompanied by a caregiver who is also willing to be recorded. If a participant agrees to this protocol, then the caregiver will read each prompt to the participant. The participant will then repeat the words spoken by the caregiver, or respond to any question asked by the caregiver. Participants also have the option to provide additional data about themselves, such as their age, race and ethnicity, and the year of their diagnosis. These "metadata tags" are completely optional but are helpful for analysis. The collected speech samples will be stored securely in a custom database built by the UIUC Beckman Institute. All samples are stored with a unique participant ID code. All samples are annotated by our UIUC research team with technical information about the acoustic waveform and other information. The entire database of speech samples will be shared with our coalition partners (Amazon, Apple, Google, Meta, and Microsoft), and, after all data collection is complete, with other universities and companies who are willing to sign our data use agreement. Each partner has signed a data use agreement with UIUC that allows these deidentified data to be used for improvements in speech recognition technology and assures the privacy of participants and confidentiality of data.