⏱️ Lectura: 10 min
When you switch conversations at a party, your brain doesn’t cut off the previous voice all at once: it keeps listening to it for a fraction of a second while it locks onto the new one. That’s what a team of researchers measured with EEG in a study published on July 16, 2026 in PLOS Biology, which documents for the first time a window of simultaneous encoding of two competing voices during a shift in auditory attention.
📑 En este artículo
The finding comes from Sara Carta, Emina Aličković, Johannes Zaar, Alejandro López Valdés, and Giovanni Di Liberto (Trinity College Dublin and collaborators), and uses Temporal Response Functions (TRF) to decode, millisecond by millisecond, which voice a person is paying attention to. The result matters for the design of smart hearing aids, brain-computer interfaces, and any system that tries to infer the focus of auditory attention from brain signals.
TL;DR
- Study published in PLOS Biology on July 16, 2026 (PLoS Biol 24(7): e3003876).
- Authors: Sara Carta, Emina Aličković, Johannes Zaar, Alejandro López Valdés, and Giovanni Di Liberto.
- Participants with normal hearing listened to two competing voices over background babble noise.
- Every 15 to 30 seconds, they were asked to switch their focus of attention between the two voices.
- EEG and Temporal Response Functions (TRF) were used to decode which voice each person was attending to.
- The cortex begins encoding the new voice before it stops encoding the previous one: there’s overlap.
- EEG alpha power drops during the transition, a sign of greater cognitive effort.
- Data, code, and audio stimuli are openly available on Zenodo in CND format.
What Happened with Auditory Attention in the Experiment
The team designed an immersive multi-talker environment: adults with normal hearing listened to two competing voices layered over background babble simulating a coffee shop or an open office. This wasn’t a sustained-attention experiment, where the listener stays locked onto a single voice the whole time: every 15 to 30 seconds, a cue told them to switch their focus to the other voice.
Using EEG, the researchers computed Temporal Response Functions (TRF) for each speech stream and confirmed they could reliably decode which of the two voices each person was attending to at any given moment. So far, nothing surprising: the cocktail party effect literature already knew that the auditory cortex distinguishes between the target voice and the background voice.
What’s new appears at the moment of the switch. Instead of a clean transition (turn off A, turn on B), the data show asymmetric disengagement and engagement: the cortex starts tracking the new target voice before it stops tracking the previous one. During that brief window, both voices are encoded simultaneously and transiently in the brain signal.
Context and History: Why Sustained Attention Dominated the Research
The neurophysiology of selective listening comes from simulated cocktail-party scenarios and uses several technologies to measure how the brain segregates the target voice from background voices: non-invasive EEG, intracranial EEG (iEEG) in patients with electrodes implanted for epilepsy, magnetoencephalography (MEG), and functional magnetic resonance imaging (fMRI). That previous work pinpointed precise areas of the auditory cortex where segregation occurs and measured substantial (though not total) suppression of linguistic processing for the ignored voice.
The problem is that almost all of that literature focused on sustained-attention tasks: the listener picks one voice and sticks with it for the entire experimental block. That left a considerable gap regarding what happens neurologically the moment someone decides to switch who they’re listening to, something that happens constantly in a real conversation. This study directly tackles that gap with a dynamic-switching paradigm every 15 to 30 seconds.
Technical Details and Performance
A TRF works like a linear regression model that maps a feature of the stimulus (for example, the acoustic envelope of a voice) against the EEG signal within a time window around each instant. If the model accurately predicts the EEG signal from Voice A’s envelope, that person is probably attending to Voice A during that stretch. Repeating this for both voices in parallel is what allows the focus of auditory attention to be decoded moment by moment.
The other key marker is alpha power (the 8-12 Hz EEG band), which drops right during the transition between voices. A drop in alpha is usually interpreted as greater cognitive effort or higher attentional demand, so the temporal pattern fits: the brain works harder precisely during the stretch where it encodes both voices at once.
flowchart TD
A["Attention on Voice A"] --> B["Cue: switch voice"]
B --> C["Simultaneous encoding of A and B"]
C --> D["Alpha power drop"]
D --> E["Voice B consolidates as target"]
E --> F["Voice A disengages"]To understand why the team chose EEG (and not another technique), it helps to compare the options available in this field:
| Method | Invasiveness | Temporal Resolution | Typical Use |
|---|---|---|---|
| EEG | Non-invasive | Milliseconds | Attention studies with many participants, like this one |
| iEEG | Invasive (implanted electrodes) | Milliseconds, high spatial precision | Patients with electrodes for epilepsy |
| MEG | Non-invasive | Milliseconds | Fine localization of cortical sources |
| fMRI | Non-invasive | Seconds | Spatial mapping of active areas, not fast transitions |
EEG is the only option in the table with both the temporal resolution (milliseconds) and the per-participant cost needed to capture an attention transition lasting fractions of a second, across dozens of subjects. That’s why it dominates this type of design.
💭 Key takeaway: beyond the acoustic signal, the team isolated cortical activity linked to lexical prediction (how well the brain anticipates the next word) and compared context-accumulation strategies built with language models (LLMs). The observed pattern suggests the listener resets their lexical context after switching attention, rather than carrying over the entire history of the previous conversation.
How to Explore the Data and Code
All the study’s material is open: the preprocessed EEG, analysis files, and code are published on Zenodo, and the audio stimuli live in the same folder under STIMULI. The data format follows the Continuous-event Neural Data (CND) standard, designed precisely for continuous EEG series aligned with stimulus events.
To reproduce a basic TRF analysis, the usual starting point in this field is the mTRF-Toolbox, which Di Liberto (a co-author of this paper) also contributed to in previous work:
# Windows, macOS, and Linux (requires Git and MATLAB or Octave)
git clone https://github.com/mickcrosse/mTRF-Toolbox.git
cd mTRF-ToolboxWith the toolbox installed, training a TRF model on a voice’s envelope and the corresponding EEG signal looks like this:
% stim: acoustic envelope of the target voice (samples x 1)
% resp: EEG signal (samples x channels)
fs = 128; % sampling frequency in Hz
tmin = -100; % minimum window in ms
tmax = 400; % maximum window in ms
lambda = 1e3; % ridge regularization
model = mTRFtrain(stim, resp, fs, 1, tmin, tmax, lambda);
[pred, stats] = mTRFpredict(stim_test, resp_test, model);
disp(stats.r) % correlation between real and predicted EEG: verifies the model decodes wellAnd to load CND-format files directly from Python before moving on to your own pipeline:
import scipy.io as sio
cnd = sio.loadmat("dataSub01.mat")
eeg = cnd["eeg"]["data"][0, 0] # samples x channels
stim = cnd["stim"]["data"][0, 0] # envelope of each voice
print(eeg.shape, stim.shape)💡 Tip: before running a full TRF across all subjects, try it with a single participant and a single voice: the Zenodo repository file has subjects separated out, so your first reproducible experiment takes minutes, not hours.
Impact and Analysis
The practical interest of this work goes beyond basic neuroscience. Hearing aids with AI-assisted beamforming and brain-computer interfaces that try to automatically focus the microphone on the voice the user wants to hear depend on decoding auditory attention in real time. If the model assumes a clean transition between voices (as most previous studies, focused on sustained attention, did), it will fail right at the most common moment in a real conversation: switching who you’re listening to.
Knowing that a window of simultaneous encoding exists, and that it comes with a measurable drop in alpha power, gives these systems an extra signal to detect that the user is “in transit” between two focuses of attention, instead of forcing a binary decision too soon. The finding about lexical context resetting is also relevant for any system that combines neural decoding with a language model: it suggests it’s not a good idea to feed the model the entire history of both voices without distinction.
As an honest limitation: the study was done with normal-hearing adults in a controlled environment, not with real hearing aid users or people with hearing loss, which is precisely the population where this kind of decoding would matter most in a product. Directly extrapolating these results to commercial hardware still requires validating them in those populations and with portable EEG hardware, which is much noisier than a lab setting.
What’s Next
The authors themselves leave open the question of exactly how the lexical context gets updated after an attention switch, beyond confirming that something like a reset happens. The logical next steps include replicating the paradigm with people with hearing loss or cochlear implant users, and reducing decoding latency enough for a hearing aid to react to the attention switch in real time, not just analyze it afterward in the lab.
📖 Summary on Telegram: View summary
Try it yourself: download the CND dataset from the Zenodo repository and run your first mTRFtrain on a single participant’s data to see the TRF curve in action today.
Frequently Asked Questions
What is the cocktail party effect?
It’s the brain’s ability to isolate a voice of interest among several simultaneous voices and background noise, like at a party with multiple conversations happening at once.
What is a Temporal Response Function (TRF)?
It’s a model that relates a feature of the stimulus, such as a voice’s acoustic envelope, to the EEG signal within a time window, and it allows predicting or decoding which stimulus a person is attending to.
Why did they use background babble noise in addition to the two voices?
So the scenario would resemble a real auditory attention situation, like an office or a bar, instead of an artificially quiet experiment.
Are the study’s data and code public?
Yes. The preprocessed EEG, analysis code, and audio stimuli are available without restrictions in the study’s Zenodo repository, in CND format.
How does this relate to smart hearing aids?
Hearing aids that try to focus the microphone on the voice the user wants to hear need to decode auditory attention in real time, and this study shows that assuming a clean transition between voices is an incorrect simplification.
What does it mean that alpha power drops during the attention switch?
The EEG alpha band (8-12 Hz) usually drops when cognitive effort increases, so its decline during the transition confirms that switching auditory focus takes active work for the brain.
References
- PLOS Biology: original article, Carta et al. 2026, “Competing speech streams are simultaneously represented in the human cortex during attention switching”.
- Zenodo: EEG dataset in CND format, audio stimuli, and the study’s analysis code.
- mTRF-Toolbox on GitHub: reference tool for training and evaluating TRF models on EEG.
- Wikipedia: general background on the cocktail party effect and selective auditory attention.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Bhautik Patel en Unsplash
0 Comments