## Can a Subscalp BCI Match Intracortical Signal Quality for Speech Decoding?

University of Melbourne spinout Fluent says the answer may be yes — at least for phrase-level communication. Using a 144-[electrode array](https://bciintel.com/glossary/electrode-array) cap placed over the motor cortex during preliminary human testing at St Vincent's Hospital Melbourne, the company demonstrated that an AI model could correctly identify a target phrase from a pool of 128 options with **96% accuracy**. The device sits beneath the scalp but outside the skull — a subscalp architecture that deliberately trades maximum signal fidelity for a dramatically lower surgical risk profile than fully intracortical systems. This positions Fluent in a distinct tier below [ECoG](https://bciintel.com/glossary/ecog) and intracortical approaches but potentially well above scalp EEG in both signal quality and practical deployability.

The preliminary testing enrolled participants who wore the cap while speaking, miming words, and imagining different phrases. Fluent co-founder Dr. Tim Mahoney reports that his research suggests signals captured beneath the skull are comparable to those captured outside it — a claim that, if independently replicated in peer-reviewed form, would meaningfully challenge the long-standing assumption that you must penetrate the skull to get clinically actionable motor cortex data for speech applications.

---

## The Signal Quality Question Is the Central One

The BCI field's conventional hierarchy runs roughly: scalp EEG at the bottom, subscalp or epidural above that, [ECoG](https://bciintel.com/glossary/ecog) (electrocorticography) placed directly on the cortical surface above that, and intracortical microelectrode arrays — such as the Utah Array used in [BrainGate Consortium](https://bciintel.com/companies/braingate) trials or the N1 chip deployed by [Neuralink Corp](https://bciintel.com/companies/neuralink) — at the top for single-unit resolution and raw throughput.

Fluent's core claim — that subscalp placement produces signals "comparable" to outside-skull placement — requires careful parsing. Mahoney is not asserting subscalp equals intracortical; he is asserting that the penalty for not penetrating the skull is smaller than previously assumed. That is a more defensible and still consequential claim. The 96% phrase identification accuracy from a 128-option pool is notable, but readers should understand what this metric does and does not represent:

- **It is a closed-set classification task**, not open-vocabulary continuous speech decoding. The model must identify one of 128 pre-defined phrases, not transcribe arbitrary speech in real time.
- The testing was **preliminary human feasibility work**, not a controlled clinical trial with peer-reviewed publication. No NCT number, trial phase designation, or published dataset has been cited in the source material.
- The electrode cap used in testing (144 electrodes, positioned externally over the motor cortex) is described as a research configuration; the implantable subscalp device architecture for the commercial product has not been detailed publicly.

This distinction matters enormously for clinical translation timelines. Early-phase feasibility data with externally worn electrode caps tells you signal processing and decoding are plausible — it does not tell you how a chronically implanted subscalp device performs across weeks, months, or years of use, nor how it handles the gliotic tissue response that complicates all implanted neural interfaces over time.

---

## What Fluent Is Actually Building

Fluent's target population is patients with motor neuron disease (including [Amyotrophic Lateral Sclerosis (ALS)](https://bciintel.com/glossary/als)) and multiple sclerosis — conditions where progressive loss of voluntary motor control eventually eliminates verbal communication. The clinical need is real and underserved: existing augmentative and alternative communication (AAC) technologies require residual motor function (eye tracking, switch access) that degrades as these diseases advance.

The device architecture as described: subscalp placement above the motor cortex, targeting the cortical representations of speech-related musculature. Mahoney frames it as capturing the motor intention signals — what he calls the "QR codes" of individual mouth and jaw movements — that persist even when the motor output pathway is compromised by disease. This is consistent with the broader [brain-computer interface](https://bciintel.com/glossary/brain-computer-interface) literature on attempted movement decoding in paralyzed patients, where motor cortex activity patterns for intended movements remain present even when downstream motor neurons cannot execute them.

The machine learning backend pairs with what Mahoney describes as context-aware large language models to correct decoding errors. This is a pragmatic engineering choice: rather than demanding near-perfect raw neural decoding accuracy, you leverage linguistic priors to error-correct in post-processing. Meta's Brain2Qwerty effort and multiple academic ECoG speech BCI groups use similar hybrid neural-language-model pipelines. Fluent's stated goal is "near-perfect speech decoding performance" when this combination is applied — a target that remains to be demonstrated in published, peer-reviewed trial data.

The company also reports building "the largest English dataset of its kind" from motor cortex brain activity, and subsequently partnering with a Japanese research team using an even larger dataset. The 128-option, 96%-accuracy result emerged from that cross-dataset collaboration. The scale of these datasets and the specific methodologies used have not been independently verified; the source material cites Mahoney's statements directly.

---

## Australia's Regulatory and Commercial Position

Mahoney explicitly identifies Australia's R&D tax incentives as a strategic asset for clinical trial execution, suggesting Fluent intends to run its pivotal trials domestically before seeking international clearance. Australia's Therapeutic Goods Administration (TGA) is recognized by most major regulatory bodies and operates under standards broadly aligned with FDA and CE Mark frameworks, which would support eventual US and European market entry.

The broader geopolitical context Mahoney references is accurate: both China and the United States are actively advancing minimally invasive BCI technologies toward commercialization. [Neuralink Corp](https://bciintel.com/companies/neuralink) is conducting its PRIME study (NCT05235853) in the US for cursor control in tetraplegia. [Synchron](https://bciintel.com/companies/synchron) has implanted its endovascular Stentrode device in patients across the US and Australia under IDE. Precision Neuroscience's Layer 7 cortical interface has received breakthrough device designation. Fluent is entering a field where the clinical bar and the competitive signal-quality bar are both rising quickly.

Fluent's differentiation argument is explicitly one of accessibility over performance ceiling: a subscalp device requires less surgical infrastructure, potentially shorter recovery, and — if the chronic performance holds — could be implanted and managed in neurology centers without dedicated neurosurgical BCI programs. For a disease like ALS where patients may have limited time and energy to invest in device management, that accessibility argument is clinically coherent. It is also the same argument being advanced, in different form, by [Synchron](https://bciintel.com/companies/synchron)'s endovascular approach — neither company is competing with Neuralink on raw bits-per-second; both are competing on risk-adjusted utility.

The motor cortex focus of Fluent's device also makes it relevant to the broader neuroprosthetics field — researchers and engineers working on robotic prosthetic limbs and humanoid robot control via neural signals will be watching whether subscalp motor cortex recordings can deliver actionable decoded intent at scale. Coverage of those intersections continues at [humanoidintel.ai](https://humanoidintel.ai).

---

## Industry Trajectory Assessment

Fluent's 96% closed-set accuracy from subscalp recordings is an encouraging early signal, but the company faces three near-term proof points before the BCI investment and clinical community will take a strong position:

1. **Chronic implant performance data.** Cap-based recordings do not predict chronic subscalp implant signal quality. The company needs longitudinal data from implanted devices.
2. **Peer-reviewed publication.** The dataset and accuracy claims need independent validation through publication. Press release science is not clinical evidence.
3. **Open-vocabulary or larger closed-set decoding.** 128 phrases is a starting point for AAC, not a full communication system. The next performance milestone should expand the vocabulary and demonstrate real-time usability in a clinical population.

If Fluent clears these bars, it would represent a genuinely useful addition to the communication BCI toolkit — not because it outperforms intracortical systems on bandwidth, but because it could reach patients who cannot access or do not want the surgical risk profile those systems require.

---

## Key Takeaways

- Fluent (University of Melbourne spinout) is developing a subscalp [brain-computer interface](https://bciintel.com/glossary/brain-computer-interface) targeting speech communication for motor neuron disease and MS patients
- Preliminary human testing used a 144-electrode cap over the motor cortex; the implantable subscalp device has not been detailed publicly
- An AI model correctly identified a target phrase from 128 options with 96% accuracy, using a dataset developed in collaboration with Japanese researchers
- This is early-stage feasibility data — no published peer-reviewed trial, no NCT registration cited in source material
- Fluent's strategy: prioritize low surgical risk and accessibility over maximum signal fidelity, using LLM-assisted error correction to close the performance gap
- Australia's R&D tax incentives are cited as a strategic asset for clinical trial execution
- The company acknowledges its device will not match intracortical BCI performance, but argues the risk-adjusted utility favors its approach for broader patient access

---

## Frequently Asked Questions

**What is Fluent's subscalp BCI and how does it differ from Neuralink?**
Fluent's device sits beneath the scalp but outside the skull, targeting the motor cortex to decode speech intentions. Neuralink's N1 implant uses intracortical microelectrodes that penetrate brain tissue, delivering higher raw signal resolution but requiring open neurosurgical implantation. Fluent trades signal ceiling for lower surgical risk.

**What does 96% accuracy from 128 phrases actually mean?**
It means an AI model, given 128 pre-defined phrase options, selected the correct one 96% of the time in preliminary testing. This is a closed-set classification result, not open-vocabulary continuous speech transcription. Real-world AAC applications would require either a larger phrase set or true continuous decoding — both are more demanding tasks.

**Has Fluent's technology been published in a peer-reviewed journal?**
The source material does not cite a peer-reviewed publication or NCT registration for Fluent's human testing. The accuracy figures and dataset claims come from statements by co-founder Dr. Tim Mahoney. Independent peer-reviewed validation has not been confirmed.

**Who are the target patients for Fluent's speech BCI?**
Primarily people with motor neuron disease (including ALS) and multiple sclerosis who have lost or are losing the ability to speak due to progressive neurological damage. These patients retain motor cortex activity for intended speech even when the output pathway is compromised.

**How does Fluent's subscalp approach compare to ECoG speech BCIs?**
ECoG (electrocorticography) places electrode grids directly on the cortical surface, requiring a craniotomy but providing higher spatial resolution than subscalp approaches. Academic groups have demonstrated online continuous speech decoding with ECoG. Fluent's subscalp architecture avoids skull penetration entirely, which reduces surgical risk but also reduces signal resolution. The company's claim that subscalp signals are "comparable" to outside-skull recordings requires peer-reviewed validation before it can be directly benchmarked against ECoG performance.

---

*Note: Results described above are from preliminary human feasibility testing and have not been reported in peer-reviewed publications or registered clinical trials as of this writing. This article does not constitute medical advice. Readers should consult clinical trial registries and published literature for regulatory and clinical decision-making.*