“Deep Learning Reinvents the Hearing Aid: Finally, Wearers of Hearing Aids Can Pick out a Voice in a Crowded Room”, 2016-12-06 (; backlinks; similar):
The human auditory system can naturally pick out a voice in a crowded room, but creating a hearing aid that mimics that ability has stumped signal processing specialists, artificial intelligence experts, and audiologists for decades. British cognitive scientist Colin Cherry first dubbed this the “cocktail party problem” in 1953.
More than six decades later, less than 25% of people who need a hearing aid actually use one…The global US $7.5$62016 billion hearing aid industry is expected to grow at 6% every year through 2020…The greatest frustration among potential users is that a hearing aid cannot distinguish between, for example, a voice and the sound of a passing car if those sounds occur at the same time. The device cranks up the volume on both, creating an incoherent din.
It’s time we solve this problem. To produce a better experience for hearing aid wearers, my lab at Ohio State University, in Columbus, recently applied machine learning based on deep neural networks to the task of segregating sounds. We have tested multiple versions of a digital filter that not only amplifies sound but can also isolate speech from background noise and automatically adjust the volumes of each separately. We believe this approach can ultimately restore a hearing-impaired person’s comprehension to match—or even exceed—that of someone with normal hearing. In fact, one of our early models boosted, 10% → 90%, the ability of some subjects to understand spoken words obscured by noise. Because it’s not necessary for listeners to understand every word in a phrase to gather its meaning, this improvement frequently meant the difference between comprehending a sentence or not…Having demonstrated promising initial results with our early classification algorithms, we decided to take the next logical step—to improve the system so it could function in noisy real-world environments, and without training for specific noises and sentences. This challenge prompted us to try to do something that had never been done before: build a machine-learning program that would run on a neural network and separate speech from noise after undergoing a sophisticated training process. The program would use the ideal binary mask to guide the training of the neural network. And it worked. In a study involving 24 test subjects, we demonstrated that this program could boost the comprehension of hearing-impaired people by about 50%.
…People in both groups showed a big improvement in their ability to comprehend sentences amid noise after the sentences were processed through our program. People with hearing impairment could decipher only 29% of words muddled by babble without the program, but they understood 84% after the processing. Several went from understanding only 10% of words in the original sample to comprehending around 90% with the program. There were similar gains for the steady-noise scenario with hearing-impaired subjects—they went 36% → 82% comprehension. Even people with normal hearing were able to better understand noisy sentences, which means our program could someday help far more people than we originally anticipated. Listeners with normal hearing understood 37% of the words spoken amid steady noise without the program, and 80% with it. For the babble, they improved from 42% of words to 78%. One of the most intriguing results of our experiment came when we asked, Could people with hearing impairment who are assisted by our program actually outperform those with normal hearing? Remarkably, the answer is yes. Listeners with hearing impairment who used our program understood nearly 20% more words in the babble and about 15% more words in steady noise than those with normal hearing who relied solely on their own auditory system to separate speech from noise. With these results, our program built from deep neural networks has come the closest to solving the cocktail party problem of any effort to date.