Here’s something I’ve been waiting for — a sound processing algorithm promising to isolate speech in noise and, once and for all, solve the “cocktail party” problem for hearing aid wearers.
A team of researchers at Ohio State University has developed a new algorithm that identifies speech-based sound waves in noisy environments and eliminates other noise, and they claim to be delivering dramatic results in tests with hearing-aid users.
The results include improvements in speech comprehension in noise by test subjects who scored only 25 percent without the new algorithm but up to 85 percent with it.
If the results hold up and the algorithm can be successfully integrated into the DSP-based sound processing systems of high-end hearing aids, the team led by Ohio State Professors Eric Healy and DeLiang “Leon” Wang may have solved the biggest problem people with hearing loss face — understanding what other people say in noisy environments.
Healy and Wang said in a news release that the team developed a unique approach using new technology to identify which sound waves are speech and which aren’t.
“Focusing on what one person is saying and ignoring the rest is something that normal-hearing listeners are very good at, and hearing-impaired listeners are very bad at,” said Healy, professor of speech and hearing science and director of Ohio State’s Speech Psychoacoustics Laboratory. “We’ve come up with a way to do the job for them, and make their limitations moot.”
Most speech processing algorithms in hearing aids identify and amplify frequencies where voices are most often located, then dampen other frequencies. That approach eliminates some of the background noise but not all of it where background chatter is at the same frequency as the voices you are trying to hear.
The Ohio State algorithm uses artificial intelligence-based machine learning to identify sound waves that are almost certainly the speech of a conversation partner standing near you. It grabs those sound waves and eliminates all the others, isolating enough of the speech to provide comprehension while dramatically reducing background noise.
“For 50 years, researchers have tried to pull out the speech from the background noise. That hasn’t worked, so we decided to try a very different approach: classify the noisy speech and retain only the parts where speech dominates the noise,” said Wang, a professor of computer science and engineering.
Why hasn’t this approach been tried before? Wang says identifying speech is only possible with very powerful, brand-new neural-network-based “deep-learning” systems that sample tens of millions of bits of real-time information, compare them, rapidly develop an understanding of their context, and instantaneously relate each bit to all other information being collected at the same time — immediately making sense of all of it in much the same way the human brain learns and comprehends.
The Ohio State team is among the first to apply neural-net-based deep-learning technology to hearing-aid design, following big successes of the new technology in other fields ranging from robotics to software development. They published results of their study in the Journal of Acoustical Society of America.
The team will be licensing its technology through Ohio State’s Technology Commercialization Office, and in the meantime the professors say they are already working with an R&D group at Starkey, one of the “Big Six” leaders in the global hearing-aid business.
In the past decade, Starkey has made a big investment in its own basic research capabilities, focusing much of its effort on the search for better algorithms that enable hearing aids to deliver comprehensible speech in noisy environments. Time will tell if the Ohio State algorithm translates into new hearing aids that solve the “cocktail party” dilemma once and for all.