Researchers tried to decode speech from brain waves. The results show why that's much harder than it looks.

Scientists built a neural network to identify spoken sounds from high-density EEG recordings and found the system works well on simple binary tasks but fails badly on real phoneme recognition, achieving error rates around 67-69%. The work is useful as a benchmark showing the actual limits of current brain-to-speech decoding, not as a working system.

This paper demolishes the hype around EEG-based speech decoding by showing what the clean benchmark results were hiding: the moment you move from toy tasks to real phoneme recognition, the system collapses. The confound problem is brutal — acoustic onset timing and electrode placement artifacts matter more than neural signal, which means most published results in this space are probably measuring artifacts, not actual speech representation in the brain. Anyone planning to invest in brain-computer interfaces for speech output should know this: the gap between 'our model works on controlled tasks' and 'our model works on speech' is not a speed bump, it's a cliff.