Researchers at the University of Illinois have a simple solution: Instead of only training the artificial intelligence that fuels it on people with near-"perfect" diction, expand the database to include people whose speech is affected by those conditions.
"Usually, what we find is that people who might have difficulty being understood in some situations will be understood just fine by people who know them well," said Mark Hasegawa-Johnson, a professor of electrical and computer engineering. "We know that it is possible to communicate with speech, it's just a matter of getting the device accustomed to your voice."
Speech recognition is used by Alexa, Siri, speech-to-text in cellphones and plenty of other technology.
It isn't uncommon for the software to make mistakes — just see the many online memes created out of flubbed speech-to-text messages — but the error rate is much higher for people with Parkinson's, for example.
"If the automatic speech recognizer is expecting the person to sound like an audiobook narrator and instead it's listening to somebody with Parkinson's, it gets almost a 20 percent error rate," Hasegawa-Johnson said. "When we train it using the speech of other people with Parkinson's, the error rate goes down to about 10 percent."
Hasegawa-Johnson remembers seeing people with cerebral palsy and certain traumatic brain injuries pioneer uses for speech recognizers, often training the technology themselves.
People with motor disabilities can benefit in big ways from speech-recognition technology: For them, talking is usually far faster and easier than typing, even with errors in the transcription.
AN EFFORT IS BORN
When Hasegawa-Johnson got to the UI in 1999, he reached out to Adrienne Perlman, head of the Department of Speech and Hearing Science at the time, to start a project for developing speech recognition for people with cerebral palsy.
They recorded 18 Urbana-area residents and created a small database.
For a while, a doctoral student used it to train speech recognizers, but it went unused for some time after he graduated.
"I did something that in hindsight was one of the most important things I did," Hasegawa-Johnson said.
He'd asked the participants if it was OK to share their recordings with researchers at other labs; 16 of the 18 said yes.
Through the 2010s, that small database became vital to projects across the world, as different labs in places like the United Kingdom and Hong Kong used it for their own projects with speech-recognition software.
"This funny thing happened: It started this kind of international competition to get the best error rates on our tiny little corpus, and that has gone on until today," Hasegawa-Johnson said.
The error rate has dropped from 70 percent to 20 percent over the last 15 years.
"But it's still an order of magnitude worse than the error rates that are being produced by commercial speech recognizers," Hasegawa-Johnson said. "The reason for that is that there's only 16 people in that database, and commercial speech recognizers are trained on thousands."
So why not create a database of thousands of people who weren't included in the original versions?
The Speech Accessibility Project is doing just that, getting as many people with Parkinson's, Down syndrome, cerebral palsy or ALS or who have suffered a speech disability due to a stroke to record themselves speaking.
NEW AVENUES
Down syndrome was a more recent addition to the list, which brought Marie Moore Channell and Laura Mattie, both associate professors of speech and hearing science, to the project.
They had already been doing other kinds of research involving people with Down syndrome, both led to the field by personal connections.
For Moore Channell, it was an older brother with Wolf-Hirschhorn syndrome, which is similar to Down syndrome but very rare.
"Growing up with him made me very aware of the needs and the barriers of accessing the community like everyone else," she said.
She wanted to do research that would make independent living more accessible for all people and started gravitating toward Down syndrome research.
"Working with this population in graduate school, I just fell in love with them," Moore Channell said. "They're a very unique community, very tight knit, very accepting and I just couldn't stop working with them."
Mattie's personal connection is a childhood friend.
"My mom did respite care and one of her families had a toddler with Down syndrome. He was one of my favorite people, even as a very young child," Mattie said. "When I got into graduate school, I realized, 'Oh, I can actually work to support these families and children in a whole different way than I've ever thought.'"
Both said they're excited for this project to increase accessibility and inclusion for the people they've been working with for so long.
"It's just this great possibility for bringing them along with us, instead of the history of people with disabilities being left behind," Mattie said. "We're making sure we're bringing them with us as we move into this more tech-savvy era."
SEEKING NEW PARTICIPANTS
Illinois residents can't participate in the project, as state data-privacy laws limit use of the data to three years and the researchers want this database to last in perpetuity, but people from most other states are welcome.
Participation in the research is actually pretty simple: Those who want to be involved sign up online, then have a phone call with a researcher.
It isn't uncommon for people to underestimate their speech capabilities, so the phone call serves as a screening process that often weeds out people who speak too clearly to qualify.
Hasegawa-Johnson said he doesn't want that to stop anybody from applying, though.
"We're trying to encourage people to just contact us and let the speech pathologist listen to them, because there's a strong tendency for people not to notice when their speech has become different from typical, because their family members can understand them just fine," he said.
Once approved, participants record audio clips reading prompts or responding to questions and submit them, all through an app.
Moore Channell and Mattie worked to make sure the prompts were accessible and relevant to people with Down syndrome — which mainly meant including prompts for things that young people would be interested in, as they were originally aimed at people with Parkinson's, who skew much older.
BIG TECH SUPPORT
Amazon, Apple, Google, Meta, and Microsoft are all supporting the Speech Accessibility Project at UI, so they'll get dibs on the database for the first six months after completion, but after that it will become more widely available.
Moore Channell said she could see this leading to things like new smart home technology that would help people with disabilities be more independent, but that remains to be seen.
For now, the goal is to make day-to-day technology use a better, more accessible experience for all.
"We talk a lot about things being accessible and inclusive, but we don't always test that out and see if it actually works, so while the technology is there, it may not do what individuals need it to," Mattie said. "Having speech technology that works and is actually picking up on their speech will open a lot of doors in their everyday life and in their jobs."
©2023 The News-Gazette (Champaign, Ill.). Distributed by Tribune Content Agency, LLC.
