Summary: For decades, researchers have asked how people can pick out a single conversation in a noisy room—a puzzle called the “cocktail party problem.” New work from MIT neuroscientists supplies a clear computational explanation for how the brain achieves selective listening.
Using a modified neural network that simulates auditory processing, the researchers show that a simple mechanism—multiplicative gain, a targeted amplification of neurons tuned to a particular voice feature like pitch or location—can account for humans’ ability to focus on one speaker. The model not only identifies the intended voice but also reproduces common human errors and spatial effects, supporting the idea that the brain acts like a feature-specific “volume knob” to direct attention.
Key Facts
- Multiplicative gain: When you attend to a voice, neural units tuned to that voice’s features increase their activity multiplicatively, boosting the target signal relative to background sounds.
- Feature cues: The model is cued with a short audio snippet of the target voice, and that cue determines which units receive increased gain—low-pitched voices amplify low-pitch units, while high-pitch units are down-weighted.
- Horizontal versus vertical separation: Both the model and human listeners are far better at separating voices that differ in left-right position than voices that differ in elevation.
- Human-like errors: The model mirrors human difficulties—such as confusing two voices of similar pitch (for example, two male voices or two female voices)—demonstrating that multiplicative gain predicts both successes and failures of selective listening.
- Potential applications: The approach could inform improvements to cochlear implant processing, helping users filter background noise and focus on a single speaker in crowded settings.
Source: MIT
How the brain isolates a single voice
When you converse in a busy environment—at a party, a crowded cafe, or a noisy workplace—your auditory system must solve a difficult problem: separate and prioritize one speech stream among many. Neurophysiological studies have long shown that neurons in auditory cortex change their responses when attention is directed to a particular sound. This new computational study demonstrates that those observed changes, implemented as multiplicative gains in a neural network, are sufficient to explain how selective listening emerges.
Josh McDermott, professor of brain and cognitive sciences at MIT and senior author of the study, explains that simply boosting the activity of processing units tuned to the attended features produces much of the behavioral profile of human auditory attention. The model reproduces a wide range of human listening behaviors without being specifically engineered to mimic human performance.
Lead author Ian Griffith, a graduate student in the Harvard Program in Speech and Hearing Biosciences and Technology working with McDermott, together with MIT graduate student R. Preston Hess, built the model and tested it across many listening conditions. Their paper appears in Nature Human Behavior.
Modeling attention with feature gains
Previous work established that neurons tuned to a target’s features—such as pitch—tend to increase firing when that target is attended. Those increases often act like a multiplicative scaling of neural responses. Until now, however, it was not proven whether such gains alone could enable selective listening in complex, real-world mixtures of voices.
The researchers adapted an existing deep neural network model of audition so each stage could apply multiplicative gains to its processing units. On each trial, the model received a short cue: an audio excerpt of the target talker. The pattern of activations evoked by that cue defined which units would receive boosting when the model encountered the mixed-audio stimulus. In effect, the cue set a feature-specific “gain profile” that emphasized units corresponding to the target voice.
For example, a cue with a low-pitched voice raises the gain for units representing low pitch and attenuates those tuned to higher pitches. The model then listened to mixtures of voices and was tasked with identifying the second word spoken by the cued talker. The researchers compared the model’s responses to human performance across a variety of conditions.
Across tests, the model’s successes and failures closely matched human listeners. It struggled in the same scenarios humans do—most notably when two competing voices shared similar pitch characteristics—supporting the hypothesis that multiplicative feature gains are a central mechanism of auditory attention.
Spatial effects and new discoveries
In addition to pitch, spatial cues play a major role in selective listening. The model naturally learned to use spatial location, and it showed improved selection when the target and distractor voices came from different horizontal positions. By exhaustively testing combinations of target and distractor positions in simulation, the researchers identified a notable horizontal advantage: left-right separation is far more effective than up-down separation. They then confirmed this asymmetry in human experiments.
McDermott highlights the value of using computational models as discovery tools: models can screen a large space of conditions quickly, revealing patterns that can then be validated with human subjects.
The team is also exploring how the same modeling approach can simulate listening through cochlear implants. That line of work aims to translate the principles of optimized feature gains into signal-processing strategies that help implant users focus on desired talkers in noisy environments.
Funding: The research was supported by the National Institutes of Health.
Key Questions Answered:
A: Listening amid many voices is a signal-to-noise problem. Multiple speech streams compete for the same neural resources. Selective attention helps by amplifying neural responses to the attended voice while reducing responses to others, effectively raising the target signal above the background.
A: Yes. When you focus on a speaker, neurons that represent that speaker’s features increase their gain, while neurons representing distractors decrease activity. The process is active suppression of competing information, not just passive ignoring.
A: The model and human data show a “horizontal advantage.” Timing and level differences between the ears provide strong cues for left-right separation, which the auditory system exploits more effectively than vertical cues. Evolutionary and environmental factors likely made horizontal localization more behaviorally relevant.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The published journal paper was reviewed in full.
- Additional explanatory context was added by the editorial staff.
About this auditory neuroscience research news
Author: Sarah McDonnell
Source: MIT
Contact: Sarah McDonnell – MIT
Image: The image is credited to Neuroscience News
Original Research: Open access. “Optimized feature gains explain and predict successes and failures of human selective listening” by Ian M. Griffith, R. Preston Hess & Josh H. McDermott. Nature Human Behavior. DOI: 10.1038/s41562-026-02414-7
Abstract
Optimized feature gains explain and predict successes and failures of human selective listening
Selective attention enables listeners to focus on particular sound sources, but it is not well understood why attention succeeds in some situations and fails in others. Neurophysiology implicates multiplicative feature gains in auditory attention, but whether such gains can predict real-world listening behavior has been unclear.
The authors optimized an artificial neural network that implements stimulus-computable feature gains to recognize a cued talker’s speech from binaural audio in complex “cocktail party” scenarios. Although not trained to replicate human behavior explicitly, the model produced human-like performance across diverse conditions, using both voice features and spatial cues. It also reproduced common selection failures and predicted new attentional effects later confirmed in human experiments. The results suggest that human-like selective listening strategies naturally emerge from optimizing feature gains for attentive listening.