Abstract
Restoring communication for people with dysarthria secondary to pontine stroke remains a critical challenge. Intracortical brain-computer interfaces (iBCIs) have demonstrated great potential for speech restoration in people with amyotrophic lateral sclerosis (ALS), with 1-24% word error rates (WERs) on a 125,000-word vocabulary. In pontine stroke, electrocorticography (ECoG) BCIs achieved 25.5% WERs with a smaller 1,024-word vocabulary. Whether intracortical BCI performance improvements extend to people with pontine stroke-induced dysarthria remains unclear. Here, we show that neural activity from a single 64-channel microelectrode array in orofacial motor cortex can predict attempted speech in a person with pontine stroke more accurately than prior ECoG BCI work and comparably to prior iBCI work. We trained a neural network decoder to predict phoneme probabilities from spiking rates and spike-band power as BrainGate2 participant ‘T16’ mimed (mouthed without vocalization) sentences from a large vocabulary. A series of language models converted these probabilities into word sequences. This decoding architecture has remained stable more than two years post-implantation, achieving a median 19.6% WER with a 125,000-word vocabulary and a median 10.0% WER with a 1,024-word vocabulary (a 60.8% reduction over prior ECoG studies). This framework also generalized beyond cue repetition, enabling T16 to communicate spontaneously via the iBCI in a question-and-answer setting with a 35.2% WER. These results demonstrate that brain-to-text decoding from a small patch of cortex can outperform ECoG-based systems in individuals with pontine stroke and is comparable to early speech iBCIs in individuals with ALS.
Introduction
Pontine (brainstem) strokes often leave individuals with "locked-in" syndrome or profound dysarthria—the inability to coordinate the muscles used for speech. While their language-processing centers remain intact, the neural "output cable" is effectively severed. Previous attempts to restore speech using Brain-Computer Interfaces (BCIs) have primarily focused on Amyotrophic Lateral Sclerosis (ALS) or utilized Electrocorticography (ECoG), which records from the brain's surface.
ECoG systems for stroke survivors have historically been limited by high error rates (around 25%) even on small vocabularies. It was previously unclear if intracortical BCIs (iBCIs)—which use penetrating microelectrodes to record individual neurons—would be effective for stroke survivors. Such strokes often lead to cortical thinning and reduced functional connectivity in the motor cortex over decades of disuse, potentially degrading the available speech signals.
Methods
Hardware and Anatomical Targeting
The study involved participant "T16," a woman who had a pontine stroke 19 years prior. The research team implanted a 3.2 x 3.2 mm 64-channel microelectrode array into the ventral premotor cortex/orofacial motor cortex (Brodmann’s area 6v). This location was selected based on anatomical markers and functional mapping to identify a "putative speech hotspot."
Decoding Architecture
The system recorded both individual neural spikes and "spike-band power" (SBP). A Recurrent Neural Network (RNN) was trained to predict the probabilities of 41 different phonemes as the participant "mimed" (mouthed without sound) sentences. To handle "neural drift" (changes in signal over time), the researchers implemented day-specific input layers. A Language Model (LM) pipeline then converted phoneme probabilities into text, utilizing a 5-gram model for real-time updates and a large 6.7-billion parameter Transformer model (OPT) to rescore and finalize predictions.
Evaluation Benchmarks
The system was evaluated using several benchmarks:
- Large Vocabulary: The Switchboard Corpus (125,000 words).
- Small Vocabulary: A 1,024-word set derived from NLTK Twitter and Cornell film corpora for comparison with ECoG studies.
- Spontaneous Communication: A set of 29 open-ended questions to test real-world conversational utility.
Results
Decoding Performance
The iBCI achieved high-accuracy decoding across all tasks. For the 125,000-word large vocabulary task, the system reached a median 19.6% word error rate (WER) at a speed of 35 words per minute. On the 1,024-word vocabulary, the WER dropped to 10.0%, representing a 60.8% reduction in error compared to the best previous ECoG-based results for pontine stroke. Crucially, the system generalized to spontaneous speech during the Q&A task, enabling communication at 27.7 words per minute with a 35.2% WER.
Signal Resilience and Stability
Despite 19 years of disuse and documented cortical thinning, the orofacial motor cortex generated highly detailed, decodable signals. While individual neurons changed their tuning day-to-day, the overall population-level structure of speech representation remained stable for approximately 40 days at a time. The model could be fine-tuned to 95% of its peak performance using only about 36 calibration sentences, requiring roughly 6 minutes of daily setup.
Discussion
These results demonstrate that high-performance decoding does not require covering vast areas of the brain. A single "patch" of cortex smaller than a fingernail provided sufficient information to outperform much larger surface-level electrode grids. This "micro-scale efficiency" suggests that intracortical recording is a viable and superior path for restoring communication in stroke survivors.
However, limitations remain. As a single-participant case study, these results require replication in a larger cohort. Furthermore, while the system performed well with mimed speech, error rates were higher in spontaneous conversation than in prompted repetition. Future work will focus on including more conversational data in training sets and exploring higher channel counts, which are predicted to drive error rates even lower following log-linear scaling trends observed in other patient populations.
Conclusion
This study proves that high-quality speech signals persist in the motor cortex decades after a brainstem stroke. By leveraging intracortical microelectrode arrays and advanced language modeling, it is possible to restore high-accuracy, large-vocabulary communication to individuals with severe dysarthria, significantly advancing the state of the art in neural prosthetics.
Summary updated