Scientists Mimic Human Ear to Improve Speech Recognition
Through extensive study of how speech is perceived by people with normal hearing, University of Wisconsin-Madison researchers have now created a method for making speech more intelligible to listeners with hearing-impairments.
Psychologist Keith Kluender and neuroscientist Rick Jenison have developed an algorithm that instead of boosting the loudness of sounds—as do most hearing aids—mimics the way the human ear works to make speech clear and recognizable.
"Our method was inspired by things we learned about how the healthy auditory system does some of its own adjustments within the sound spectrum," says Kluender. "Biology came up with this solution first. We've just put it into an algorithm."
Along with other acoustic characteristics, differences between speech sounds are signaled by peaks and valleys in energy across the sound spectrum, explains Kluender. Healthy ears distinguish sounds from one another by tracking changes in these peaks and valleys over time, with processes in the brain operating to enhance these changes. Hearing impairment, on the other hand, "tends to smear the differences between sound elements," says Kluender, making the peaks flatter and the valleys shallower. As a result, sounds blur together, and speech becomes muddy-sounding and difficult to understand. Finding ways to make sounds more distinct, rather than simply louder, poses a major challenge to hearing aid designers, says Kluender.
He likens sound amplification alone to turning up the brightness on a computer monitor washed out by lack of contrast; increased brightness by itself doesn't make the screen's contents more readable, and may in fact distort them further.
Kluender and Jenison's approach instead aims to digitally enhance the contrast between the elements of sound. Their algorithm carves the incoming auditory signal into a large number of spectral channels, or bands; sharpens the differences between sound elements within the bands; and then recombines them to produce a clearer output signal. Through extensive analyses of speech, they have also tuned the algorithm so it accentuates those components of the signal containing useful speech information, such as consonant sounds.
"You begin with speech," says Kluender, "because what a hearing-impaired person would like most from a hearing aid is the ability to communicate." Jenison has just completed a simulation program, allowing listeners to hear the algorithm's effects on speech that has been digitally altered to sound as it would to a hearing-impaired individual.
Kluender and Jenison's technology is available for licensing through the Wisconsin Alumni Research Foundation, the patent management organization of the UW-Madison.