Google building impaired speech dataset for speech recognition inclusivity
Google has partnered with the ALS Therapy Development Institute and the ALS Residence Initiative (ALSRI) to gather voice samples from people with the neurodegenerative disease for its Project Euphonia, Forbes reports.
Project Euphonia was launched by the company at its recent I/O Conference, with the goal of improving speech recognition for people who are deaf or have neurological conditions including ALS, Parkinson’s, stroke, multiple sclerosis or traumatic brain injury.
“You can turn on your lights, your music or communicate with someone. But this only works if the technology can actually recognize your voice and transcribe it,” Google AI Product Manager Julie Cattiau told Forbes.
Cattiau says its currently an open question if a machine learning model can be trained to understand different people with ALS, let alone other conditions. Models may need to be trained specifically for individuals. Although her team has collected tens of thousands of recorded phrases, it will require hundreds of thousands more.
“Impaired speech is a very difficult data set to put together. It’s not as simple as asking people to record phrases, and there’s no data set just lying around,” Cattiau says. “We have to first put it together, and that’s a lot of work.”
The Project Euphonia team is also attempting to train AI systems to interpret gestures, vocalizations, and facial expressions to trigger actions, according to the report.
accuracy | biometrics | dataset | Google | machine learning | speech recognition | voice recognition