Deep learning technology could help you filter out unwanted background noises while keeping the ones you like.
ADVERTISEMENTCould artificial intelligence (AI) help you filter the noise of babies crying or sirens while keeping the singing of the birds?
A team from the University of Washington believe they’ve figured out how to be able to zone out on background noises at will.
They have developed algorithms using deep-learning technology that allow users to choose the sounds they can listen to when they’re wearing their headphones. They have called this system "semantic hearing".
"At a high level, we use noise-cancelling headphones to suppress all sounds in the environment. But we run on a neural network on the smartphone to extract the sounds of interest and in real-time play it back into the ear through the headphones," Shyam Gollakota, professor of Computer Science and Engineering, told Euronews Next.
So, how does it work? Headphones send recorded sounds to a connected smartphone, blocking other noises. Users can pick sounds from 20 categories - like birds chirping - using voice commands or an app, with only the chosen sounds being played.
"The target sounds we play through the speakers would have to be consistent with what the wearer sees in the environment. For this reason, we only have 20 milliseconds to process the input sound, extract target sounds and play them back on the headphones," said Bandhav Veluri, PhD student in Computer Science, explaining that they had to create "a very efficient deep learning model".
This time constraint explains why the system relies on smartphones rather than cloud servers.
'Ripe for being launch into consumer market'
"What is interesting here is that when people typically talk about neural networks and artificial intelligence these days they're familiar with large language models like ChatGPT," said Gollakota.
"This requires very large models that run in huge data centres which is really not possible for our application. We designed a special neural network that can run on a smartphone".
Tested in places like offices, streets, and parks, the system could pick out sirens, bird sounds, alarms, and specific noises while getting rid of all other background sounds.
When 22 people gave feedback on the system's sound output, they generally said it sounded better than the original recording, according to the study’s results.
However, in some situations, the system found it hard to tell the difference between sounds that are quite similar, like singing and talking. The researchers suggest that training the models with more real-world data might help improve these results.
The system "is ripe for being launched into the consumer market," said Veluri.
"These two trends [noise-cancelling technology and deep learning] present opportunities for creating the future of intelligent wearables, with real-world capabilities that so far have been in the realm of science fiction," said Gollakota.
"I am very excited that this is the right time to create these intelligent headsets and that we will start seeing intelligence in our headsets within the next five years," he added.
