EEG Doesn’t Read Your Mind: Speech Decoding Without the Hype

EEG-to-Text Is Still Bottlenecked by the Signal, Not the Model

The simple truth: non-invasive EEG speech decoding still isn't freeform "thoughts to text." The Frontiers review on covert speech decoding explicitly lists low SNR, muscle artifacts, poor spatial resolution, and practical setup issues; the 2025 review adds inter-subject variability and small, heterogeneous datasets.

So take demos with a grain of salt. A 2024 Frontiers article reports around 70-80% accuracy for non-invasive EEG on tightly constrained tasks, but on more complex scenarios the numbers drop to about 20-50%. This isn't a "bad product"—it's the physics of the channel: electrodes, gel, muscle artifacts, session variability, individual differences.

My first question for any such announcement: what exactly is being decoded? A free phrase, a selected word, a symbol from an interface, a scene class, or a reaction to a stimulus? These are often bundled into one flashy headline, but engineering-wise they’re different tasks with different error costs.

Real Pipelines Are Usually Smarter Than the Headlines

The most reliable path today isn't "reading the mind" but narrowing the choice space. Eye-tracking shows where the person is looking, ERPs like P300 confirm the reaction to a target symbol or word, and the model classifies not from an infinite language but from a small menu.

This is where marketing often starts juggling words. If a system displays a keyboard, catches gaze over a symbol, and confirms the choice via an event-related potential, that can be a useful BCI. But calling it full-blown thoughts-to-text is a stretch.

Video decoding follows a similar story. Over the decade, it's not brain-decoding magic that improved dramatically, but the generator quality: GANs, diffusion, DiT, flow matching. The pipeline "signal → scene classification → fancy generation" may look like reading visual experience, though EEG often carries coarse semantics, not an exact frame.

Where This Actually Shifts the Game

Real value lives where the task is constrained by design: command selection, simple games, button interfaces, clinical communication, imaginary movements. There, low bandwidth doesn't kill the scenario if the system is honestly designed around the limits.

The biggest risk is cross-patient accuracy. A model that looks decent on one person or one session can collapse on another patient, another day, or a different electrode setup. Invasive approaches capture much better signals, but even there the pipeline complexity quickly brings everyone back to earth.

So my dry take: EEG BCI is alive and can be very useful, but "thought decoding" is mostly still packaging for a classifier, a selection interface, and a good generator. The real intrigue isn't wow-demos—it's the honest boundary between signal and embellishment.

We've already explored cutting-edge AI tools for transcribing and summarizing meetings, like Otter.ai and Granola. Their accuracy and hallucination risks resonate with the challenge of converting EEG signals into meaningful text, where data interpretation is even more complex.