Neural dynamics of natural speech: Microscale analysis from long-term intracortical recordings

Neural representations of speech have been extensively characterized in controlled settings, yet their stability over extended timescales remains untested. We address this gap using continuous week-long intracortical recordings from a Utah microelectrode array in the right inferior parietal lobule of an epileptic patient during unconstrained natural behavior. This naturalistic paradigm required tailored computational approaches: fine-tuned Swiss-German speech recognition, encoding and decoding pipelines across multiple linguistic levels, and rigorous validation through synthetic spike train experiments. Decoding speech presence reached 62% and language identity 40-50% (for 50% and 33% chance levels, respectively), while stimulus-locked responses, word onsets and semantic content, remained at chance. Crucially, classifiers failed to generalize across time, and neural trajectories wandered continuously through state space rather than clustering by condition. These findings expose a tension between laboratory paradigms and ecological reality. Representations that appear stable in controlled sessions may be snapshots of drifting dynamics invisible to short experiments. We discuss implications for speech BCI development, where long-term stability cannot be assumed.