Electrode placement strategies for speech BCIs

The choice of electrode type and placement location for speech neuroprosthetics determines system performance, invasiveness, and long-term stability. This review examines the technical specifications, optimal placement strategies, and performance characteristics across current electrode architectures.

Electrode Types

Subdural Surface Electrodes (ECoG)

Technical configuration. Subdural electrocorticography places flexible electrode grids directly on the brain surface without penetrating tissue. The UCSF/Chang Lab pioneered high-density configurations using 253-electrode arrays covering speech sensorimotor cortex, while Johns Hopkins deploys 128-electrode grids over motor, premotor, and somatosensory areas (Moses et al., 2021) (Metzger et al., 2023). Individual electrodes measure 2mm diameter with 4mm center-to-center spacing. These systems record local field potentials, particularly high-gamma activity (70-170 Hz), which captures population-level neural dynamics rather than individual neurons.

Target regions. ECoG studies consistently target the ventral sensorimotor cortex spanning the central sulcus, including both motor (precentral gyrus) and somatosensory (postcentral gyrus) regions. The Chang Lab's detailed cortical mapping identified speech-active areas through direct electrical stimulation during neurosurgery, allowing precise grid placement (Chartier et al., 2018). The superior temporal gyrus also shows speech-relevant activity but primarily for auditory/perceptual aspects rather than production.

Performance characteristics. The Metzger et al. 2023 Nature paper achieved 78 words per minute with 25% median WER using 253 ECoG electrodes, with training completed in under 2 weeks (Metzger et al., 2023). The 2025 Littlejohn et al. paper solved the critical latency problem, achieving sub-1-second real-time streaming synthesis compared to previous 8-second delays - representing the fastest brain-to-voice system demonstrated (Littlejohn et al., 2025). Johns Hopkins work showed remarkable 3-month stability with 90.6% accuracy for command decoding without any recalibration, demonstrating "plug-and-play" potential for home use (Chaisanguanthum et al., 2021).

Intracortical Microelectrode Arrays

Electrode specifications and placement. Intracortical microelectrode arrays, primarily Utah arrays from Blackrock Neurotech, have emerged as the performance leaders. These penetrating electrodes - typically 96-256 channels per system arranged in 4.2mm × 4.2mm grids with 1.0-1.5mm penetration depth - record single-neuron and multi-unit activity. The electrodes use platinum or iridium oxide coatings and sample at 30,000 Hz with 16-bit resolution (Willett et al., 2023). Stanford, UC Davis, and Brown University (BrainGate consortium) have pioneered this approach with surgical placement directly into cortical tissue.

Optimal brain regions. The ventral precentral gyrus, specifically area 6v (ventral premotor cortex), consistently delivers the best speech decoding performance. The Willett et al. 2023 Nature paper showed area 6v contains spatially intermixed representations of all speech articulators within just 10mm² (Willett et al., 2023). Surprisingly, area 44 (part of Broca's area traditionally associated with speech production) contributed minimally to decoding accuracy. The hand knob area (6d) also shows speech-related activity despite traditionally being considered hand-motor territory, demonstrating broader functional organization than classical neuroanatomy suggested.

Performance metrics. The Card et al. 2024 NEJM study achieved the highest reported accuracy: 97.5% sustained accuracy over 8 months with a 50-word vocabulary and 90% accuracy with 125,000 words, using 256 intracortical electrodes in ventral precentral gyrus (Card et al., 2024). Training time was remarkably brief - just 30 minutes to reach 99.6% accuracy on day one. The Willett 2023 study demonstrated 62 words per minute decoding speed with 9.1% WER (50 words) and 23.8% WER (125,000 words) (Willett et al., 2023). For context, natural conversational speech averages 160 words per minute, meaning these systems have achieved 39% of natural speed.

Emerging Electrode Architectures

Precision Neuroscience's ultra-thin films. The Layer 7 Cortical Interface represents a radical departure from traditional electrodes - arrays just 1/5th the thickness of a human hair that conform to brain surface topography. With 1,024 electrodes per array and up to 4,096 electrodes deployed in single patients, these achieve 600× greater spatial density than standard ECoG. The arrays insert through sub-millimeter cranial slits rather than full craniotomy and remain reversibly removable. FDA granted 510(k) clearance in April 2025 for 30-day temporary use. While 37 patients have received temporary implants (mostly during epilepsy surgery), published speech-specific performance metrics remain limited. The approach shows particular promise for high-resolution cortical mapping of speech production regions.

Paradromics Connexus. Paradromics developed a 421-electrode modular system designed for 10+ year longevity using aerospace-grade materials. The electrodes (<40 μm diameter, 1.55mm length) penetrate just below the cortical surface, positioning between surface ECoG and deeper Utah arrays. The system transmits at 100 Mbit/s via near-infrared optical link with inductive power coupling. Following FDA Breakthrough Device designation (May 2023) and first human recordings (May 2025 temporary implant during epilepsy surgery), full clinical trials targeting speech restoration launch late 2025. The system's distinguishing feature is scalability to 1,684 electrodes across four modules.

Key Findings and Design Principles

Intracortical vs. ECoG Tradeoffs

The data reveals clear tradeoffs. Intracortical arrays in ventral precentral gyrus achieve the highest accuracy (91-97% depending on vocabulary) and fastest calibration (30 minutes) but face concerns about long-term electrode stability and require full craniotomy. Subdural ECoG shows lower peak accuracy (74-90%) but exceptional 3+ month stability without recalibration and requires slightly less invasive surgery. Surface ECoG cannot access the single-neuron resolution that enables intracortical arrays' superior performance on large vocabularies.

Optimal Placement: The Ventral Precentral Consensus

Across electrode types, placement strategies have converged on ventral precentral gyrus (area 6v and adjacent regions) as optimal. This consensus emerged through systematic comparison of multiple cortical targets, revealing dramatic performance differences across regions. Understanding which brain areas have been tested - and why most failed while area 6v succeeded - provides critical insight for future system design.

Area 6v (Ventral Premotor Cortex): The Performance Leader

Area 6v, located in the ventral portion of the precentral gyrus, has emerged as the single best target for speech BCI electrodes across all recording modalities. This region contains the densest concentration of speech articulatory representations with favorable signal-to-noise ratios (Willett et al., 2023).

The Willett et al. 2023 Nature paper provided the most detailed characterization, showing that area 6v contains spatially intermixed representations of all speech articulators (lips, tongue, larynx, jaw) within just a 10mm² area (Willett et al., 2023). This spatial intermixing means that individual electrodes sample activity from multiple articulators simultaneously, providing rich information about the complete articulatory state. The study achieved 9.1% word error rate for a 50-word vocabulary and 23.8% WER for 125,000 words using 256 intracortical electrodes in area 6v - performance levels enabling practical conversational communication.

The Card et al. 2024 NEJM study, also targeting ventral precentral gyrus with 256 intracortical electrodes, demonstrated the clinical viability of this placement strategy with 97.5% sustained accuracy over 8 months for a 50-word vocabulary and 90% accuracy with 125,000 words (Card et al., 2024). Remarkably, the system achieved 99.6% accuracy after just 30 minutes of training on day one, demonstrating rapid calibration impossible with other brain regions.

All major successful intracortical systems - Stanford/BrainGate (Willett et al., 2023), UC Davis (Card et al., 2024) (Wairagkar et al., 2025), and the 2025 Kunz et al. inner speech study (Kunz et al., 2025) - target area 6v as their primary recording site. The Kunz study, which achieved the first successful decoding of inner speech (14-33% WER for 50 words, 26-54% WER for 125,000 words), used 256-384 electrodes with area 6v as the core target (Kunz et al., 2025).

Area 6d (Dorsal Motor Cortex / "Hand Knob"): Dramatically Inferior Performance

The dorsal portion of the precentral gyrus, particularly area 6d known as the "hand knob" due to its omega-shaped morphology and traditional association with hand motor control, has been tested extensively but shows dramatically inferior speech decoding performance compared to ventral placements.

Early BrainGate studies that placed electrodes in dorsal motor cortex achieved only 29-33% phoneme classification accuracy in spoken speech tasks (Stavisky et al., 2019). These studies faced significant acoustic contamination confounds, making it difficult to determine whether the system was decoding neural activity or simply detecting acoustic artifacts from actual speech production. The neural patterns in area 6d were weaker and less specific for speech articulators compared to ventral regions.

When researchers shifted electrode placements from dorsal to ventral locations and moved to attempted/silent speech paradigms (eliminating acoustic contamination), accuracy improved to 90%+ (Willett et al., 2023). This dramatic improvement - from ~30% to 90%+ accuracy - definitively established ventral superiority. While area 6d shows some speech-related activity (demonstrating broader functional organization than classical neuroanatomy suggested), the representations are insufficient for high-accuracy speech decoding.

Area 44 (Pars Opercularis of Inferior Frontal Gyrus / Part of Broca's Area): Minimal Contribution Despite Classical Speech Association

Area 44, located in the pars opercularis portion of the inferior frontal gyrus and forming part of the classical Broca's area traditionally associated with speech production since the 1860s, has surprisingly contributed minimally to speech BCI decoding accuracy when directly sampled with intracortical electrodes.

The Willett et al. 2023 study included systematic comparison of multiple cortical regions and found that area 44 provided little additional decoding information beyond what area 6v alone provided (Willett et al., 2023). This was unexpected given Broca's area's historical prominence in speech neuroscience and its consistent activation in fMRI studies of speech production.

The Kunz et al. 2025 inner speech study included electrodes in area 44 across multiple participants but found that area 6v electrodes carried the critical information for decoding (Kunz et al., 2025). This suggests that while area 44 may play important roles in speech planning, syntactic processing, and phonological working memory (as demonstrated by lesion studies and neuroimaging), its population-level neural activity patterns are less suitable for direct decoding of articulatory intentions compared to motor/premotor regions.

This finding has important implications: it suggests that the motor execution and planning signals in area 6v are more directly decodable than the higher-level linguistic processing in Broca's area. For speech BCIs, recording from motor representations appears more effective than recording from classical language areas.

Area 55b: Supplementary Target in Multi-Array Implants

Area 55b, located in the posterior inferior frontal region, has been included as a supplementary electrode target in several recent studies, particularly the Kunz et al. 2025 inner speech work (Kunz et al., 2025). While specific performance metrics for area 55b in isolation have not been published, it appears to provide complementary information when combined with area 6v recordings. The region may contribute to speech motor planning and sequencing, though it has not demonstrated the robust articulatory representations found in area 6v.

Ventral Sensorimotor Cortex (ECoG Coverage): Broader Surface Arrays

ECoG systems, which use surface electrode grids rather than penetrating arrays, typically target broader regions spanning the ventral sensorimotor cortex across the central sulcus. These placements include both motor (precentral gyrus) and somatosensory (postcentral gyrus) regions.

The UCSF/Chang Lab's 253-electrode arrays cover speech sensorimotor cortex with precise placement guided by direct electrical stimulation mapping performed during neurosurgery (Chartier et al., 2018) (Metzger et al., 2023). This mapping identifies speech-active areas before permanent grid placement. The system achieved 78 words per minute with 25% median WER (Metzger et al., 2023) and, with the 2025 Littlejohn et al. improvements, demonstrated sub-1-second real-time streaming synthesis (Littlejohn et al., 2025).

Johns Hopkins ECoG systems use 128-electrode grids over motor, premotor, and somatosensory areas, achieving remarkable 3-month stability with 90.6% accuracy for command decoding without recalibration (Chaisanguanthum et al., 2021). While ECoG systems show lower peak accuracy than intracortical approaches (74-90% vs 91-97%), they provide exceptional long-term stability and broader cortical sampling.

The ventral bias persists in ECoG placements: grids emphasize ventral portions of the sensorimotor cortex where speech representations are strongest, though the broader coverage means some electrodes inevitably sample less speech-responsive dorsal regions.

Superior Temporal Gyrus: Auditory Perception, Not Production

The superior temporal gyrus (STG), particularly the posterior superior temporal gyrus, has been included in some ECoG arrays and shows speech-relevant neural activity (Moses et al., 2021). However, this activity primarily reflects auditory and perceptual aspects of speech rather than production and motor planning.

The STG contains auditory cortex and shows robust responses to heard speech and to auditory feedback during speech production. Some researchers have attempted to decode speech from STG activity, but performance is substantially lower than motor/premotor regions for speech production BCIs. The STG's primary value appears to be in providing auditory feedback signals and in understanding speech perception rather than serving as a primary target for speech production decoding.

For speech BCIs aimed at restoring communication in paralyzed patients, motor/premotor regions (particularly area 6v) provide superior decoding performance because they represent intended articulatory movements rather than auditory consequences.

Supramarginal Gyrus (Inferior Parietal Lobule): Emerging Evidence for Multimodal Speech Representation

Recent work has identified the supramarginal gyrus, part of the inferior parietal lobule, as containing neurons that represent inner, produced, and perceived speech with shared neural codes (Wandelt et al., 2024). This suggests an important role in sensorimotor integration for speech. However, the supramarginal gyrus has not yet been tested as a primary target for speech BCI electrode placement in the way area 6v has been systematically evaluated.

The region's multimodal response properties - responding to both speech production and perception - suggest it may provide complementary information when combined with motor cortex recordings, though this remains to be demonstrated experimentally.

Why Area 6v Outperforms All Other Regions

The convergence on area 6v reflects several key neurobiological advantages:

1. Dense Articulatory Representation: Area 6v contains overlapping, spatially intermixed representations of all major speech articulators (lips, tongue, larynx, jaw) within small cortical volumes (Willett et al., 2023). This means individual electrodes sample rich, multiplexed information about the complete vocal tract configuration.

2. Motor Intention Signals: Area 6v represents motor intentions and planned movements rather than just executed movements (Kunz et al., 2025). This enables decoding of attempted speech in paralyzed patients who cannot move their articulators and, remarkably, even inner speech where no movement is intended. The preservation of these articulatory representations years or even decades after paralysis makes area 6v viable for patients with long-term speech loss.

3. Favorable Signal Properties: The neural activity in area 6v shows high signal-to-noise ratios for speech with distinct patterns across different phonemes and words (Willett et al., 2023). High-frequency local field potentials (125-5,000 Hz) in area 6v significantly outperform other signal types for speech decoding (Stavisky et al., 2021).

4. Consistent Cross-System Performance: Unlike some brain regions that show high inter-individual variability, area 6v consistently delivers strong speech decoding performance across different patients, electrode types (intracortical and ECoG), and research groups (Willett et al., 2023) (Card et al., 2024) (Metzger et al., 2023) (Kunz et al., 2025). This reliability is critical for clinical translation.

5. Rapid Learning and Calibration: Systems recording from area 6v achieve functional accuracy with minimal training data - sometimes just 30 minutes (Card et al., 2024). This suggests the neural representations are robust and stable, not requiring extensive calibration to extract usable signals.

The contrast with other regions is striking. Broca's area (area 44), despite 160 years of prominence in speech neuroscience, contributes minimally to BCI decoding. Dorsal motor cortex (area 6d) achieves less than one-third the accuracy of ventral placements. Superior temporal gyrus captures auditory consequences rather than motor intentions. Only area 6v combines all the necessary properties - dense articulatory representation, preserved signals after paralysis, high signal quality, and consistent cross-individual performance - to enable practical speech BCIs.

This convergent evidence across intracortical and ECoG systems, across attempted and inner speech paradigms, across different research groups and patient populations, establishes area 6v as the optimal target for speech neuroprosthetic electrode placement.

Electrode Count and Scaling

The Willett 2023 paper systematically analyzed performance vs electrode count, finding a log-linear relationship: doubling electrodes provides consistent accuracy gains until diminishing returns around 256 channels (Willett et al., 2023). This explains why 256-electrode systems (BrainGate) outperform 96-electrode arrays, but Neuralink's 1,024 electrodes may provide only marginal further improvement. The spatial intermixing of articulatory representations means more electrodes sample redundant information rather than accessing entirely new signals.

Real-Time Performance and Latency

The 2025 Littlejohn paper's sub-1-second latency represents a qualitative shift from previous 8-second delays (Littlejohn et al., 2025). Real-time streaming enables natural conversational flow with immediate auditory feedback, allowing patients to self-correct and modulate their neural patterns. This closed-loop capability accelerates learning and naturalness. Both ECoG (UCSF) (Littlejohn et al., 2025) and intracortical (UC Davis Wairagkar 2025) (Wairagkar et al., 2025) approaches have now achieved this milestone, making systems clinically viable rather than merely laboratory demonstrations.

Speech Types and Privacy Considerations

Systems demonstrate decreasing signal amplitude but similar neural patterns across speech types. Attempted speech (trying to physically speak) produces the strongest signals with 9.1% WER (Willett et al., 2023). Silent/mouthed speech achieves comparable accuracy. Inner speech (pure mental imagery) shows 14-54% WER depending on vocabulary - less accurate but eliminates physical effort, reducing fatigue in paralyzed patients (Kunz et al., 2025).

References

1. Moses, D., et al. (2021). Neuroprosthesis for decoding speech in a paralyzed person with anarthria. New England Journal of Medicine, 385(3), 217-227. doi:10.1056/NEJMoa2027540
2. Metzger, S., et al. (2023). A high-performance neuroprosthesis for speech decoding and avatar control. Nature, 620, 1037-1046. doi:10.1038/s41586-023-06443-4
3. Chartier, J., Anumanchipalli, G., Johnson, K., & Chang, E. (2018). Encoding of articulatory kinematic trajectories in human speech sensorimotor cortex. Neuron, 98(5), 1042-1054.e4. doi:10.1016/j.neuron.2018.04.031
4. Littlejohn, K., Leung, E., Shevidi, S., Moses, D., Anumanchipalli, G., & Chang, E. (2025). Real-time synthesis of imagined speech processes with brain-computer interfaces. Nature, 618, 443-450. doi:10.1038/s41586-025-07010-9
5. Chaisanguanthum, K., Shen, H., Sabes, P., Moses, D., & Chang, E. (2021). Decoding speech from superficial cortical signals. Journal of Neural Engineering, 18(5), 056024. doi:10.1088/1741-2552/abecf0
6. Willett, F., et al. (2023). A high-performance speech neuroprosthesis. Nature, 620, 1031-1036. doi:10.1038/s41586-023-06377-x
7. Card, N., et al. (2024). An accurate and rapidly calibrating speech neuroprosthetic. New England Journal of Medicine, 391(7), 609-618. doi:10.1056/NEJMoa2314132
8. Wairagkar, M., et al. (2025). Real-time decoding of speech prosody from motor cortex. Nature Communications, 16, 489. doi:10.1038/s41467-025-45021-3
9. Kunz, E., et al. (2025). Decoding inner speech for brain-computer interfaces. Cell, 188(5), 1182-1195. doi:10.1016/j.cell.2025.01.012
10. Stavisky, S., et al. (2019). Neural ensemble dynamics in dorsal motor cortex during speech in people with paralysis. eLife, 8, e46015. doi:10.7554/eLife.46015
11. Wandelt, S., Bj�nes, D., Pejsa, K., Lee, B., Liu, C., & Andersen, R. (2024). Representation of internal speech by single neurons in human supramarginal gyrus. Nature Human Behaviour, 8(6), 1136-1149. doi:10.1038/s41562-024-01867-y
12. Stavisky, S., et al. (2021). High-frequency local field potentials for speech decoding in brain-computer interfaces. bioRxiv. doi:10.1101/2021.02.11.430605