Give Yourself the Silent Treatment

With laptops, smartphones, and wireless internet connections readily available to us, it is now quite easy to get work done wherever we happen to be. A coffee shop, park, or airport is every bit as appropriate as a traditional office space for many people. But aside from having the right tools for work and communication, there are a few other considerations that arise when working in public spaces, with background noise being chief among them.

The constant chatter and clatter can be exceedingly distracting, but popping in earplugs to drown it out completely is not really a good plan since there are other sounds that we do need to be aware of. Phone notifications and people that are either nearby or joining on a video call must be heard. What we really need is something like a sound bubble that allows us to hear only the sounds that are close to us, while filtering out everything that is more distant.

That may sound a bit like something out of a science fiction story, but a team led by engineers at the University of Washington has made it a reality. They have developed a prototype system that leverages a pair of headphones and artificial intelligence (AI) to allow users to set a programmable radius around them in which they can hear the sounds. Everything else is muffled to the point that it is practically inaudible.

The proof of concept device is built on top of a commercial pair of Sony WH-1000XM4 headphones. These headphones have a noise-canceling feature, which does part of the job automatically — with all sound blocked, it leaves the team the job of reintroducing the sounds that should be heard. A Seeed Studio ReSpeaker six-channel microphone array was attached to the headband to capture environmental sounds, and both a Raspberry Pi 4 Model B and an Orange Pi 5B were evaluated as potential processing units.

Next, an AI model was built and trained to estimate the distance a sound is from the microphones by analyzing the time difference between when different components of the sound arrive at each microphone in the array. No suitable dataset existed for the purpose of training this model, so the team put their headphones on a rotating mannequin head and created their own labeled dataset.

Once trained, the algorithm ran directly on the single-board computers and analyzed all incoming sounds. If the predicted distance from which a sound came was within the programmable sound bubble, it was played through the headphone speakers. If not, it was simply ignored and filtered out by the noise-canceling feature. Testing revealed that this system was able to operate fast enough to facilitate real-time communication, with the Raspberry Pi slightly edging out the Orange Pi in terms of performance.

The researchers would like to ultimately commercialize their technology, but there is some work to be done before that can happen. For starters, noise-canceling headphones do allow some residual noise through. They plan to install an inner microphone to monitor for these sounds so that they can be masked. Furthermore, the work has only focused on indoor environments so far, so at some point the team will have to turn their attention to the additional complexities of outdoor environments to make the headphones more robust.