AI Could Diagnose and Help People With Speech Conditions—Here's How

Recognition should work with unusual patterns, too

  • The tech industry is partnering with researchers to help AI understand people with disabilities or unusual speech patterns. 
  • Speech recognition has become a part of everyday life, but the tech doesn’t always recognize certain speech patterns. 
  • Improved speech recognition could help with employment and education.

A person in a wheelchair using a smartphone.

RyanJLane / Getty Images

Artificial intelligence (AI) could soon offer more help to those with speech disabilities. 

Big tech companies are partnering with the University of Illinois to form the Speech Accessibility Project to upgrade AI's understanding of people with disabilities or unusual speech patterns. The project will gather a set of high-quality, diverse speech samples that will help improve speech technologies.

"Being able to devise new interventions and screening tools will help us be more proactive in early detection of conditions in children and help us customize more specific therapies for a patient's condition," Karen Panetta, a professor of electrical and computer engineering at Tufts University and an IEEE Fellow, who is not involved in the project, told Lifewire in an email interview.

The Power of Speech

Speech recognition, found in many software programs and voice assistants, has become a part of many people's everyday lives. However, experts say these systems don't always recognize certain speech patterns, particularly those associated with disabilities. 

Companies are now going back and adding more inclusive examples to their data sets.

The Speech Accessibility Project will work to make speech technologies more inclusive by creating a dataset of representative speech samples, according to the blog post about the project. The new project will collect samples from people with all types of speech disabilities (Lou Gehrig's disease (ALS), Parkinson's, cerebral palsy, Down syndrome, and others) to ensure that they are included in the program. The information can train machine learning models to better understand a range of speech patterns.

"For example, if the original dataset only included a single iteration of 'wash,' then someone from the Pacific Northwest who might pronounce the word as 'war-sh' would not be able to interact with the program properly," Matthew Luken, the vice president of Deque Systems, a company that supports digital accessibility initiatives told Lifewire in an email interview. "Both pronunciations need to be in the dataset for comparison."

Luken said software is often built for the median population, meaning it tends to leave people out from the very first steps of the development process, with edge cases added to future versions. "Voice technology took this approach—originally designed with a narrow set of examples, which did not include people with disabilities," Luken added. "Companies are now going back and adding more inclusive examples to their data sets. This means that in future versions of the software, people with a stutter will have a higher chance of communicating with the software."

A person in a wheel chair wearing a headset and using a touch screen computer.

Horsche / Getty Images

An Urgent Need

Many people with disabilities, particularly people with physical, sensory, and cognitive disabilities, rely on technologies that incorporate AI with speech recognition to access the world around them, Katy Schmid, the director of Education and Technology at The Arc, an organization for people with intellectual and developmental disabilities, told Lifewire via email. 

Schmid pointed out that 85 percent of people with disabilities are unemployed. She said, "There are still access barriers that exist when it comes to using technology that can leverage a person with disabilities [or] speech differences to receive the proper accommodations they need for meaningful employment.”

Students with disabilities "could much more easily complete their schoolwork and advance their education and potential post-secondary and college opportunities if AI [and] speech recognition is created so that it can recognize all speech, not just speech that is considered to be clear or 'normal,'” Schmid said. 

Panetta said that AI is already being used to learn speech patterns and facial expression recognition to help develop new interventions for children with autism. AI can take in disparate forms of data, like eye tracking, to see what someone is looking at, what they are saying, and what expressions they have—even mouth positions. 

Cognitive psychologists can help to look at all these kinds of data to establish training sets for AI to establish what are good speech patterns versus those that are not, Panetta said. “Then, using this information, the AI can help therapists identify specific positions and patterns and guide the patient with interventions and therapy,” she added.

For Panetta, the research is personal as well as professional. She said that as a parent who has a child with hearing loss and speech challenges, "it’s important to me that my child be able to fully participate in society and flourish in all the activities he enjoys, including education and his passion for music.”

Was this page helpful?