Can simple inner speech utterances be decoded from EEG reliably enough for brain-computer interfaces, and what limits classification accuracy?

How good is inner-speech decoding right now?

The honest answer is: better than guessing, but not good enough to trust. With four possible words, pure chance is 25%. A 2021 study using a convolutional neural network on eight subjects reached 29.7% average accuracy for four internally spoken words — only slightly above chance [4]. A later controlled benchmark on ten subjects compared four network designs on the same public dataset and found the best, a hybrid model, at 28.8% [5]. That hybrid was the only model in the comparison whose accuracy was statistically significantly above the 25% chance level; the three standard baselines did not clear that bar [5].

Not every study lands in the same range. Using inter-trial coherence — a measure of how consistently the phase of the EEG signal lines up across repeated trials of the same word — one group reported average accuracies of 56.08% with a nearest-neighbor classifier and 59.55% with a support vector machine, across four words, ten subjects, and three sessions each [1]. The authors describe this as comparable to or better than earlier work rather than as a solved problem, and they note that existing methods generally fall short of what real-life use would require [1]. The gap between the ~30% and ~60% figures is itself informative: results depend heavily on the words chosen, the feature used, and the evaluation setup, so a single headline number does not describe the field.

Why the brain signal is so hard to read

EEG is recorded from electrodes on the scalp, so it picks up the summed electrical activity of huge numbers of neurons through bone and skin. The neural trace of silently saying a word is faint against that background, and the signal is non-stationary — its statistical properties drift over time, even within a session [5]. Phase-based work names the same obstacles directly: low signal-to-noise ratio, complex variability between subjects, and the difficulty of isolating distinct neural patterns tied to specific imagined words [3].

A useful comparison is trying to hear one person's whisper in a crowded room using a microphone across the street. The whisper is there, but so is everything else. One response is to stop looking at how large the signal is and look instead at how consistently it repeats: inter-trial coherence exploits the fact that the same imagined word may produce a repeatable phase pattern across trials even when the raw amplitude is buried in noise [1]. Phase-based features are argued to be more robust to amplitude fluctuations and noise interference for the same reason [3]. These are plausible strategies, and both report improvements over traditional approaches, but neither removes the underlying weakness of the measurement.

What actually moves the accuracy up or down

The single biggest factor appears to be whether the decoder is tuned to the individual. In the game-navigation study, a model trained on other people's data and tested on entirely new subjects reached only 45% accuracy for four directional commands [2]. Fine-tuning with just half of the new subject's own data raised that to 88% [2]. That jump is the clearest illustration in this evidence of how much person-specific calibration matters — and it also shows that the 45% figure is a cross-subject result, not a general measure of how well inner speech can be decoded.

Model architecture, by contrast, seems to matter much less than one might hope. In the controlled benchmark, the hybrid model beat the standard designs on average, but not by a statistically significant margin against any single one of them, and each individual branch performed at roughly its baseline level of about 26% [5]. With only ten subjects, that study could only detect differences of about four percentage points or larger, so the null comparisons are inconclusive rather than proof that the architectures are equivalent [5]. The authors also ran the whole protocol twice on identical hardware: the baselines reproduced exactly, but the hybrid's mean shifted by two tenths of a point, enough to move one comparison from significant to non-significant [5]. Small accuracy differences at this scale are not stable enough to build design choices on.

What would it take before this is usable?

For someone who has lost the ability to speak or move, a decoder that reads intended words directly would be far more natural than the indirect mental tasks most BCIs rely on [5]. But the numbers above describe four-word tasks. Reported four-class accuracies in the literature rarely climb much past the low thirties, and the benchmark's own results fell within a prior range of roughly 26–36% on that dataset [5]. Scaling from four words to a usable vocabulary, while keeping error rates low enough that a misread command is not dangerous or exhausting, is a much harder problem than the four-class results suggest.

The most promising near-term path in this evidence is not a better generic model but personalization: collecting some data from the individual user and adapting the decoder to them, as the fine-tuning result demonstrates [2]. The open question is how much calibration data that requires and whether the gain holds outside a single game task. Until cross-subject accuracy improves and the person-to-person variability described in phase-based work is addressed [3], inner speech is a promising but not yet dependable control signal.

Sources used in this answer

Related research questions