Objective: 1. Obtain high-fidelity, speech data from 400 people with Parkinson’s (PWP) to train automatic speech recognition (ASR) systems.

2. Develop and implement procedures to recruit, screen, train and monitor PWP during speech tasks for analysis in this project.

Background: Notably, 90% of 8 million PWP have speech disorders. These disorders limit access to everyday voice activated devices (mobile phones, smart speakers) as well as commonly used virtual receptionists, all who use ASR systems which were trained on non-disordered speech. This presents challenges for PWP with disordered speech to communicate using ASR[1].

This Speech Accessibility Project (SAP) is designed to improve ASR for PWP by retraining models on disordered speech.

Method: Recruiting PWP was the first step. From our prior experience (e.g. Project Euphonia [2]), attrition and screening failure rates were 60% for mild to moderate speakers. SAP [1] targeted moderate to severe/profound speakers, and the attrition and screen failure rate increased to 72%.

The recruiting database was established through LSVT Global relationships with Davis Phinney Foundation, other PD organizations, and via flyers, podcasts, social media, presentations, word of mouth (see Figure 1 for 123 contacts) [3].

A team of 6 Speech Mentors, certified speech clinicians with expertise in PD and technology, managed PWP through the course of the project.

Screening ensured high-fidelity data.  PWP were screened on the following variables: significant speech disorder, appropriate technology, cognitive or motor challenges too profound to proceed and/or potential use of home support.

Training PWP occurred remotely on the SAP recording platform, ensuring accuracy in data input.

Mentors provided oversight throughout the generation of 450 prompts/per PWP (~5 hours of recording time).

1,388 PWP and Parkinsonism were screened for the study to acquire high-fidelity speech from 400 PWP to train ASR systems.

Results: Initial data using disordered speech to train ASR reported Word Error Rate (WER) improvement from 36.3% to 23.7%[1]. More recently, Microsoft reported accuracy gains in speech recognition ranging from 18-60% in their Azure AI app[4].

Conclusion: Initial positive gains observed here strongly motivate continued efforts to improve ASR for PWP. Speech Mentors were key in obtaining project data.

This project was supported by Apple, Amazon, Google, Microsoft and Meta

SAP Referral Sources LSVT Global [figure 1]

References: 1. Hasegawa-Johnson, M., Zheng, X., Kim, H., Mendes, C., Dickinson, M., Hege, E., Zwilling, C., Channell, M. M., Mattie, L., Hodges, H., Ramig, L., Bellard, M., Shebanek, M., Sarι, L., Kalgaonkar, K., Frerichs, D., Bigham, J. P., Findlater, L., Lea, C., … MacDonald, B. (2024). Community-Supported Shared Infrastructure in Support of Speech Accessibility. Journal of Speech, Language, and Hearing Research, 67(11), 4162-4175.
https://doi.org/10.1044/2024_JSLHR-24-00122

2. MacDonald, R. L., Jiang, P.-P., Cattiau, J., Heywood, R., Cave, R., Seaver, K., Ladewig, M., Tobin, J., Brenner, M. P., Nelson, P. C., Green, J. R., & Tomanek, K. (2021). Disordered speech data collection: Lessons learned at 1 million utterances from Project Euphonia. In Proceedings of Interspeech 2021 (pp. 4833-4837).
https://doi.org/10.21437/Interspeech.2021-697

3. Zwilling, C. E., Hasegawa-Johnson, M., Mendes, C., Kim, H., Channell, M. M., Mattie, L., Barkhimer, A., Dickinson, M., Ramig, L., Hodges, H., Bradshaw, A., & Carter, S. (2025). The Speech Accessibility Project: Best Practices for Collection and Curation of Dysarthric Speech. Manuscript submitted for presentation. This paper has been submitted to Interspeech 2025, to be held in Rotterdam, The Netherlands, August 17-22, 2025.

4. Dickson, Meg, (2025). Speech Accessibility Project data leads to recognition improvements on Microsoft Azure. Speech Accessibility Project, Beckman Institute for Advanced Science and Technology, University of Illinois Campaign-Urbana.

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/improving-automatic-speech-recognition-for-speakers-with-parkinsons-disease/