Cogan Lab

doi.orgShared latent representations of speech production for cross-patient speech decoding - Nature CommunicationsCurrent speech brain-computer interfaces (BCIs) rely on patient-specific decoding approaches. Here, the authors show that patient-specific data can be aligned to a shared space that preserves speech i...

Can a speech BCI learn across people? In Nature Communications, led by @zspald.bsky.social: High-density µECoG reveals shared neural dynamics that can be aligned to improve cross-patient speech decoding, even with little data from a new participant. doi.org/10.1038/s414... 1/5

4/5

Thank you to all the patients who participated in our work, our colleagues in the Viventi lab, and our neurosurgeon collaborators! @dukebrain.bsky.social @dukemedschool.bsky.social @dukeneurosurgery.bsky.social 5/5

2/5

3/5

doi.orgDistinct neural processes link speech planning and execution - Nature Human BehaviourDirect recordings from the human brain reveal a hierarchy in planning to speak, with neural activity organizing whole syllables before the individual sounds that compose them are sequenced.

How does a speech plan become fluent movement? In Nature Human Behaviour, led by Kumar Duraivel: Intracranial recordings reveal how hierarchical speech plans become continuous motor sequences across planning, articulation, and monitoring networks. doi.org/10.1038/s415... 1/5

A big thank you to all the patients for participating in our research, all our co-authors including the Viventi Lab, and our neurosurgical team @dukebrain.bsky.social @gregoryhickok.bsky.social @dukemedschool.bsky.social @dukeneurosurgery.bsky.social @dukeubme.bsky.social 5/5

nature.comHow speech plans become spoken sequences - Nature Human BehaviourSpeaking requires the brain to rapidly convert planned speech into precisely ordered vocal movements. Intracranial recordings reveal a cascade of neural processes in which syllable-level plans in the ...

Nature Human Behaviour also published a Research Briefing on the work, with a short overview of the findings, check it out! www.nature.com/articles/s41...

3/5

4/5

2/5

17 Aug

Two new papers from the lab in Nature Human Behaviour and Nature Communications, spanning the neural mechanisms of speech production and new approaches to speech neurotechnology! 🧠 Speech plans → fluent motor sequences 🗣️ Speech BCIs across people Links below 👇 @dukebrain.bsky.social 1/3

17 Aug

Distinct neural processes link speech planning and execution Nature Human Behaviour doi.org/10.1038/s415... 2/3

17 Aug

Shared latent representations of speech production for cross-patient speech decoding Nature Communications doi.org/10.1038/s414... 3/3

Last week, Areti Majumdar (2nd year PhD student) presented @jerrytang.bsky.social and colleague’s seminal paper on reconstructing continuous language from non-invasive brain recordings. This 🧵 explores our thoughts and questions.

❔3⃣: The decoder was trained on speech perception and successfully generalized to imagined speech, suggesting that it captures a shared semantic representation across tasks. How well would you expect this type of model to generalize to overt speech production?

Thank you to the authors @ucberkeleyofficial.bsky.social / @utaustin.bsky.social for your work, and we look forward to following its progress! cc: @shaileejain.bsky.social @amandalebel.bsky.social @alexanderhuth.bsky.social

❔2⃣: Looking closely at the qualitative examples for imagined speech, even the general gist seems pretty far off from the target transcripts in some cases.

❔2⃣ (con't): While this is kind of expected from a non-invasive decoder trained purely on speech perception data, what is the actual floor for reconstruction accuracy we should be looking for in a viable semantic decoder?

❔1⃣ : To test if the decoder could be consciously resisted, the authors showed that subjects silently performing cognitive tasks lowered decoding performance relative to passive listening.

❔1⃣ (con't): However, could this drop in accuracy instead reflect a shift in attention away from the background story and toward the cognitive task? If so, was there any evidence that the decoder outputs reflected the content or demands of the cognitive task itself?

🤍3⃣: The finding of highly redundant semantic representations across multiple, distinct cortical networks was very interesting.

🤍3⃣ (con't): The ability to independently decode the same narrative content from regions such as prefrontal cortex and association cortex has important implications for the resilience and flexibility of future semantic BCIs.

🤍1⃣: The integration of a generative language model prior with the encoding model was a very clever way to solve the inverse problem.

🤍2⃣: The ability of the decoder to generalize across multiple task conditions is very compelling and shows the robustness of the underlying semantic representations.

Cogan Lab reposted

Tobias Egner

@tobiasegner.bsky.social

New paper out: Led by super grad student Jim Zhang and in collaboration with Greg Cogan, we review and synthesize Insights into the neural mechanisms supporting cognitive stability and flexibility from human intracranial electroencephalography. For free access: authors.elsevier.com/a/1nRYWY3M3p...

Our new review paper lead by CNAP PhD Student Jim Zhang and @tobiasegner.bsky.social is now published in Neuroscience & Biobehavioral Reviews: “Insights into the neural mechanisms supporting cognitive stability and flexibility from human intracranial EEG.” doi.org/10.1016/j.ne... 1/5

doi.orgRedirecting

The framework suggests medial frontal theta helps broadcast control demands, beta regulates whether task sets are maintained or updated, and neural populations separate shared from task-specific information. 4/5

authors.elsevier.comPlease wait whilst we redirect youAll content on this site: Copyright © 2026 Elsevier B.V., its licensors, and contributors. All rights are reserved, including those for text and data mining, AI training, and similar technologies. For all open access content, the relevant licensing terms apply.

Check it out here: authors.elsevier.com/a/1nRYWY3M3p... This link is free before 9/3/2026 @dukebrain.bsky.social 5/5

doi.orgRedirecting

We synthesize human iEEG studies of conflict control and task switching, two core components of cognitive control, that let us stay focused when needed, yet rapidly update goals when circumstances change. 2/5

We propose two interacting mechanisms at two different levels: theta and beta oscillations coordinate when and where control signals are routed, while high-gamma activity and population coding reveal local engagement and what task-relevant information is represented. 3/5

Last week, Nanlin Shi (postdoctoral associate) presented T. Lei et al.’s new Nature paper, “Active dissociation of intracortical spiking and high gamma activity,” on whether intracortical HGA reflects local spiking or broader population activity.

❔2⃣: Since HGA and spike rate remain moderately coupled during hand control, does the study reject the spike-proxy interpretation, or mainly show that the coupling can be actively broken during neurofeedback; and how should we interpret this under spontaneous behavior?

❔3⃣: How should this conclusion be interpreted for ECoG, where each electrode may average over a similar or larger spatial scale than the entire Utah array? Thank you to @caraido.bsky.social, @joshuaiglaser.bsky.social and the Slutzky Lab for your work, and we look forward its progress!

🤍2⃣: Spike-triggered averaging provides an elegant temporal analysis of the relationship between spiking and HGA. 🤍3⃣: The paper refines the interpretation of HGA by linking it to distributed population activity rather than only nearby spiking.

❔1⃣: Under what conditions is HGA a reliable proxy for spiking, and under what conditions does this interpretation break down?

nature.comActive dissociation of intracortical spiking and high gamma activity - NatureA brain–machine interface is used in monkeys to investigate the biophysical underpinnings of cortical high gamma-band activity, a signal that is often studied in the context of many brain functions.

This thread explores our thoughts (likes & questions) nature.com/articles/s41... 🤍1⃣: The neurofeedback BMI paradigm offers a strong causal test of whether HGA and local spike rate can be dissociated.

Last week, Zac Spalding (@zspald.bsky.social‬, 4th year PhD student, @dukeubme.bsky.social‬) presented Aditya Singh and colleagues’ 2025 paper on multi-regional, multi-subject neural speech decoding using from SEEG electrodes.

❔3️⃣: In the multi-subject REO analysis that showed robustness to loss of SMC electrodes (Fig. 4b), did you investigate which non-SMC regions contributed the most in the absence of SMC electrodes? Could this robustness be driven by auditory feedback signals when SMC electrodes are removed?

Thank you to the authors at the UTHealth Tandon Lab for your work! cc: Tessy Thomas, Jinglong Li, @gregoryhickok.bsky.social, Xaq Pitkow, Nitin Tandon

❔1️⃣ (con’t): For example, maybe models with training sets that more widely sample distributed speech network areas are more robust to regional exclusion, while training sets composed of focal coverage in a singular region may not be as generalizable.

❔2️⃣: There is variability in PER gains as the transfer learning is applied between top-performing training subjects and inference subjects (Fig. 5c). Is this variability again driven by the amount of coverage correlation between training and inference subjects (as in Fig. 3e) or something else?

🤍3️⃣: The cross-task transfer learning analysis (tongue-twister to TIMIT) is important for understanding changes in transfer learning performance as task structure changes.

❔1️⃣ : In multi-subject models, is there any effect of electrode coverage across training patients on PER in inference patients?

🤍1️⃣: The regional exclusion analysis is an interesting way to investigate robustness to loss of information from important cortical areas.

🤍2️⃣: The comparison of multiple network transfer strategies (full, readout, recurrent) provides useful insight into necessary model architectures to capture shared speech representations.