Community-Supported Shared Infrastructure in Support of Speech Accessibility.

Area of Science:

• Speech Technology
• Machine Learning
• Disability Studies

Background:

• The Speech Accessibility Project (SAP) was established to address the need for diverse speech data in machine learning.
• Individuals with speech disabilities often face challenges with current automatic speech recognition (ASR) systems.
• Accessible speech technology is crucial for inclusivity and equitable access to digital tools.

Purpose of the Study:

• To introduce the Speech Accessibility Project (SAP) as a valuable resource for researchers in ASR and machine learning.
• To provide a baseline analysis of the first released data package from the SAP.
• To facilitate the development of improved ASR systems for individuals with speech disabilities.

Main Methods:

• Collected and curated U.S. English speech data from individuals with speech and/or language disabilities.
• Participants recorded speech remotely via a web portal, using personal devices and assistive technology as needed.
• Speech data were manually transcribed, with a subset annotated for diagnostic patterns; participants were assigned to training, development, and test sets for ASR experiments.

Main Results:

• The initial data package includes speech from 211 individuals with dysarthria related to Parkinson's disease.
• Baseline ASR achieved a 36.3% word error rate on test speech from individuals with dysarthria, compared to 3.4% for typical speakers.
• Fine-tuning the ASR model reduced the word error rate to 23.7% for the dysarthric speech data.

Conclusions:

• A substantial corpus of dysarthric and dysphonic speech can significantly advance speech technology for people with disabilities.
• The SAP data distribution is expected to accelerate research and development in accessible speech technology.
• Continued research using SAP data holds the potential to create more equitable and effective ASR solutions.