Intracortical speech brain–computer interfaces
(BCIs) convert motor cortical activity into text, but the relation-
ship between representational quality and end-to-end decoding
accuracy is poorly studied. This work presents a multi-level
analysis of a pretrained GRU-based decoder on the Brain-to-Text
’25 benchmark (10,948 trials, 45 sessions, 20 months). Linear
probing of GRU embeddings results in 84.04% phoneme clas-
sification accuracy, substantially exceeding the 28.98% baseline
on raw neural features, suggesting that the encoder learns rich
phonemic structure that the decoder does not fully utilize. A
3-level decomposition — phoneme probe accuracy, in-sequence
phoneme accuracy, and full sequence accuracy — attributes
the resulting decoding gap to contextual decoding costs and
insertion errors. The full system with a 5-gram Weighted Finite
State Transducer (WFST) language model achieves 4.34% word
error rate with 81.0% of sentences perfectly decoded on a
held-out set of 1,718 sentence-level trials. Over this 20-month
span — to our knowledge the longest continuous longitudinal
evaluation reported for a single intracortical speech BCI partic-
ipant — performance declined significantly, by 11.3 percentage
points, highlighting the practical challenge of brain-signal drift.
Together, these results show three systematic mismatches —
representational-to-decoding, phoneme-to-word, and longitudinal
drift — and motivate the need for context-aware representations,
word-level language modeling, and continual session adaptation
for robust long-term BCI deployment.

Brain–computer interface, intracortical speech decoding, sentence decoding, word error rate, GRU, CTC, di- phone modeling, longitudinal stability