Solution to 'cocktail party problem' could help people with hearing loss

Have you ever struggled to pick out your friend's voice over other conversations in a crowded room? Scientists call this challenge the "cocktail party problem," and it can be especially difficult for people with hearing loss.

Most hearing aids come with directional filters that help users focus on sounds in front of them. They're best at reducing static background noise, but falter in more complex acoustic scenarios, such as when the user is among cocktail-party guests who are standing close together and speaking at a similar volume.

Now, a new algorithm could improve how hearing aids tackle the cocktail party problem. The model, dubbed the "biologically oriented sound segregation algorithm" (BOSSA), draws inspiration from the brain's auditory system, which uses inputs from both ears to locate the source of a noise and can filter out sound by location.

Alexander Boyd, a doctoral student in biomedical engineering at Boston University, compared directional filters and BOSSA to flashlights, in that they highlight what is in their path.

Related: 'Vestigial' human ear-wiggling muscle actually flexes when we're straining to hear

"BOSSA is a new flashlight that has a tighter beam that's more selective," he told Live Science. Compared with the standard filters, BOSSA should be better at distinguishing between speakers — though it still needs to be tested in real-world scenarios with proper hearing aids.

Boyd led a recent lab test of BOSSA, whose results were published April 22 in the journal Communications Engineering. In the experiment, participants with hearing loss donned headphones playing audio designed to simulate five people speaking simultaneously and from different angles around the listener.

The audio was filtered through either BOSSA or a more traditional hearing-aid algorithm, and the participants compared both filters to how they heard the audio without additional processing.

In each trial, participants were asked to follow sentences spoken by one of the five speakers. The volume of the "target speaker" relative to the other speakers varied between trials. When the target speaker was standing within 30 degrees of the listener in either direction, the participants could make out a greater proportion of words at a lower volume threshold with BOSSA than with the conventional algorithm or when unassisted.

The conventional algorithm did seem to serve users better than BOSSA in distinguishing speech from static noise. However, this was tested in only four of the eight participants.

The standard algorithm works by reducing distracting sounds by boosting the signal-to-noise ratio for sounds coming from a given direction. By comparison, BOSSA transforms sound waves into spikes of input that the algorithm can process, similar to how the cochlea in the inner ear converts vibrations from sound waves into signals transmitted by neurons.

The algorithm emulates how special cells in the midbrain — the uppermost portion of the brainstem that connects the brain and spinal cord — respond selectively to sounds coming from a given direction. These spatially tuned cells judge direction based on differences in the timing and volume of sound inputs to each ear.

Boyd said this aspect of BOSSA drew from studies of the midbrain in barn owls, which have sophisticated spatial sensing abilities since they rely on sound cues to locate prey. The BOSSA-filtered signals are then reconstructed into sound for the listener.

BOSSA is modeled on the nervous system's "bottom up" attention pathway, which gathers bits of sensory information that are then interpreted by the brain. These sensory inputs govern which aspects of the environment warrant focus and which can be ignored.

Related: Our outer ears may have come from ancient fish gills, scientists discover

But attention is also dictated by a "top down" pathway, in which a person's prior knowledge and current goals shape their perception. In this case, an individual can decide what is relevant to focus on. These two modes of processing aren't necessarily mutually exclusive; for instance, your friend's voice might jump out at you both because you recognize it and because they're shouting over the sound of a crowd.

BOSSA's "bottom-up" approach can help people focus on speech coming from a predetermined location, but in real life, people rapidly shift their attention to different conversations. "You can't do that with this algorithm," said Michael Stone, an audiology researcher at the University of Manchester in the U.K. who was not involved in the new study.

Stone added that the study didn't replicate how sounds echo and reverberate in real life, especially in indoor settings. Still, he said BOSSA could be more practical for hearing aids than algorithms based on deep neural networks, another emerging approach to sound filtering.

Deep neural network models need extensive training to be prepared for all the different configurations of speakers the user may encounter. And once implemented, the computational demands of these models require a lot of power. BOSSA is simpler by comparison, relying mainly on the spatial difference between two sounds.

BOSSA may also be more transparent than the "black box" of deep neural networks, said Fan-Gang Zeng, professor of otolaryngology at University of California, Irvine, who was not involved with the research. That means it would be easier to interpret how sound inputs become algorithmic outputs, perhaps making the model simpler to refine.

Zeng added that BOSSA may require further refining as it is studied in more-realistic scenarios. The researchers plan to test BOSSA in proper hearing aids, rather than in headphones, and also hope to develop a steering mechanism to help users direct the algorithm's focus.