High-Throughput Neural and Predictive Assistive Communication

This direction is split into other directions: information transfer rate bottleneck. Its last version stays readable and citable here.

Version 1 of 169 papers80 conceptsfrozen 2026-10-04

The fundamental challenge in assistive technology is a "bandwidth bottleneck." While a typical person speaks at roughly 150 words per minute (wpm), individuals with severe motor impairments—such as those with ALS or total paralysis—often communicate at rates below 10 wpm using eye-gaze trackers or brain-computer interfaces (BCIs). This disparity creates a "silence gap" that prevents natural social interaction.

To bridge this gap, researchers are attacking the problem from two sides: increasing the raw data rate of neural interfaces (the hardware path) and using artificial intelligence to expand sparse inputs into rich language (the software path).

The Information Transfer Rate Bottleneck

In the field of BCI, the standard metric for performance is the Information Transfer Rate (ITR). As Peter Meinicke and Helge J. Ritter noted in Improving Transfer Rates in Brain Computer Interfacing: A Case Study, traditional systems are often slowed down by the need for repetitions to ensure accuracy. If a system requires a user to focus on a stimulus multiple times to confirm a selection, the effective bandwidth drops significantly.

José del R. Millán highlights that non-invasive interfaces, like EEG, are particularly constrained by low ITRs, often struggling to exceed 2 bits per second. This is barely enough to move a cursor, let alone facilitate a conversation. To solve this, researchers like Guido Dornhege and Gabriel Curio have proposed moving from binary "left vs. right" selections to multi-class systems, finding a "sweet spot" that maximizes ITR by increasing the number of available commands without overwhelming the user's cognitive load.

Scaling Neural Decoding

The most direct way to increase throughput is to record higher-fidelity signals. While non-invasive methods are safer, they are "lossy." Thomas Navin Lal and Thilo Hinterberger demonstrated that invasive Electrocorticography (ECoG) can provide the high-speed, reliable communication channels that locked-in patients require.

Recent breakthroughs have pushed these limits further. In A high-performance speech neuroprosthesis, Erin M. Kunz and colleagues achieved a decoding rate of 62 wpm by recording from single neurons in the premotor cortex. Similarly, Sean L. Metzger and David A. Moses reached 78 wpm in A high-performance neuroprosthesis for speech decoding and avatar control. These systems work by decoding the intent to speak directly into text and synthesized audio, bypassing the slow process of character-by-character typing.

Intelligence as a Bandwidth Multiplier

If the physical "pipe" for data is narrow, the alternative is to send more "compressed" information. This is the core of modern Augmentative and Alternative Communication (AAC). Instead of typing every letter, the user provides a "hint," and a Large Language Model (LLM) expands it.

Shanqing Cai, Subhashini Venugopalan, and Katrin Tomanek developed SpeakFaster, a system that expands word-initial abbreviations (e.g., "iipitb" for "it is pretty in the backyard"). As detailed in Context-Aware Abbreviation Expansion Using Large Language Models, this approach achieved keystroke savings of up to 77%. By using conversational context, the LLM can predict the most likely full sentence from a tiny fraction of the characters.

This "soft-prompting" or "keyword-steering" approach is also seen in the work of Ramesh Manuvinakurike and Lama Nachman. Their Controllable Response Generation system allows users to steer AI-generated responses using simple cues, drastically reducing the physical and mental exhaustion associated with eye-gaze typing.

Hybrid Paradigms and Collaborative Control

The future of the field lies in "hybrid" systems that combine multiple signal types or interaction modes. Adrien Combaz demonstrated that simultaneously detecting P300 (attention-based) and SSVEP (flicker-based) brain signals can increase the number of available commands without increasing the time required for selection.

Furthermore, the concept of "shared control" is emerging. In systems like ACAT-G by Xueyuan Lu and Saurav Sahay, the AI isn't just a passive tool; it’s an active collaborator. The system generates potential responses, and the user provides minimal feedback to guide the dialogue. This reduces the "motor cost" per idea, allowing the user to maintain agency while the AI handles the heavy lifting of linguistic construction.

By combining high-fidelity neural decoding with the predictive power of LLMs, the research direction is moving toward a future where the "communication gap" is closed not just by faster hardware, but by smarter interfaces that understand human intent from minimal cues.

Papers in this version

Showing 60 of 69 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_high_throughput_neural_and_predictive_as_75cfd542,
  title = {High-Throughput Neural and Predictive Assistive Communication},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-throughput-neural-and-predictive-assistive-communication/rdr_75cfd54270c246c0a858274f4ec6eb0e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_high_throughput_neural_and_predictive_as_75cfd542_v1,
  title = {High-Throughput Neural and Predictive Assistive Communication},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-throughput-neural-and-predictive-assistive-communication/rdr_75cfd54270c246c0a858274f4ec6eb0e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}