Brain Encodes Two Speakers at Once During Attention Switch

A study published July 16, 2026 in PLOS Biology reveals that when listeners switch attention between two competing speakers, the brain briefly encodes both speech streams simultaneously. The finding challenges the long-held assumption that attention switching is a sequential process of disengaging from one stream then engaging with another.

Researchers from Trinity College Dublin and collaborators recorded EEG from normal-hearing adults in an immersive multi-talker environment. Participants listened to two TED talk streams from front speakers while background babble played from rear speakers. Every 15–30 seconds, a visual cue instructed them to switch attention to the other stream.

Temporal Response Functions Reveal Overlap

Using Temporal Response Functions (TRFs) to model neural tracking of each speech stream, the team found that engagement with the new target stream began before disengagement from the previous target was complete. This created a brief period—on the order of hundreds of milliseconds—where both streams were simultaneously represented in the cortex.

> "Our results indicate asymmetric disengagement and engagement processes during attention switches, where the neural tracking of the new target stream emerges before disengaging from the previous target," the authors write.

The transition was mirrored by a reduction in EEG alpha power, which the authors interpret as a marker of cognitive effort during the switch.

Lexical Context Reset via LLMs

The study went beyond neural tracking to investigate how the brain updates lexical predictions after a switch. They compared four context-accumulation strategies constructed using Large Language Models (LLMs) to model how listeners rebuild their understanding of the new target stream.

Findings point to a "reset" in lexical context after switching attention—the brain discards the previous speaker's word predictions and starts fresh with the new speaker's context. This suggests that the cost of switching attention includes rebuilding the predictive model of speech.

Implications for Brain-Computer Interfaces and Hearing Aids

For developers working on brain-computer interfaces (BCIs) or next-generation hearing aids, these results provide a neural timeline of attention switching. Current attention-decoding algorithms often assume sustained attention; incorporating the transient overlap could improve real-time switch detection.

The EEG dataset and analysis code are publicly available on Zenodo (https://zenodo.org/records/20569817) in the Continuous-event Neural Data (CND) format, enabling replication and extension.

Methodology Details

  • 16 normal-hearing adults participated
  • Two front speakers at ±30° azimuth, four rear speakers for babble
  • Switch cues every 15–30 seconds, trials lasted 3 minutes
  • TRF models with encoding window lengths from 0 to 500 ms
  • Attention decoding accuracy was significant for all window lengths (p < 0.05)

Why It Matters for AI and Audio Processing

The finding that the brain simultaneously tracks two streams has parallels in machine learning models for source separation and attention mechanisms. Current transformer-based models typically attend to one source at a time; this biological evidence suggests that a brief multi-source representation might improve switching performance.

The use of LLMs to model lexical context accumulation also offers a new way to evaluate how well language models capture human-like prediction updating—a potential benchmark for next-generation speech AI.