# Does Mamba Beat GRU for Intracortical Brain-to-Text Decoding?

The answer, as of today's preprint, is no — and the margin matters for anyone betting on architectural modernization to unlock faster, more accurate speech [brain-computer interface](https://bciintel.com/glossary/brain-computer-interface) systems.

In a controlled study published July 30, 2026 on arXiv (2607.26751), researchers Lucas Zamora Vera and Jose A. Gonzalez-Lopez ran a systematic 2×2 comparison on the public Brain-to-Text '25 benchmark: GRU versus hybrid Mamba decoder, crossed with phoneme versus character output targets, all trained under a CTC (Connectionist Temporal Classification) objective. Their headline result — a phoneme-level GRU achieves **12.62% phoneme error rate (PER) and 21.19% word error rate (WER)** — holds against the Mamba hybrid, which is described as competitive but unable to surpass the recurrent baseline. On the character side, the best GRU reaches **13.39% character error rate (CER) and 26.28% WER** after language model rescoring.

This is a preprint from a single research group using a public benchmark. Results have not been independently replicated on live implant data, and the study does not involve human participants directly — its findings apply to the neural signal datasets underpinning the Brain-to-Text '25 competition.

---

## What the 2×2 Architecture Comparison Actually Tested

The field's current best-practice pipeline for intracortical speech decoding pairs a neural-sequence decoder with an external language model for rescoring. The decoder converts neural population activity — typically from Utah arrays or similar intracortical electrode arrays recording in motor speech cortex — into a sequence of phonemes or characters, which the language model then disambiguates into words.

Two questions have gone largely unanswered until now:

1. **Does Mamba, the selective state-space model architecture that has challenged transformers in sequence modeling, improve on GRU-based recurrent decoders for this task?**
2. **Does the choice of output granularity — phonemes versus characters — interact meaningfully with decoder architecture?**

Zamora Vera and Gonzalez-Lopez address both simultaneously. The four conditions — phoneme-GRU, character-GRU, phoneme-Mamba hybrid, character-Mamba hybrid — were trained under one reproducible protocol using a CTC loss, allowing direct attribution of performance differences to architecture and output target rather than training variability.

The recurrent baseline wins. The phoneme-GRU combination produces the lowest PER and WER. The Mamba hybrid is described as competitive, but the authors are explicit: it does not surpass the GRU.

---

## Error Analysis: Representation-Dependent Failure Modes

The paper's ablations and error analysis are arguably more instructive than the headline numbers. The authors identify what they call **representation-dependent failures** — error patterns that differ qualitatively depending on whether the model is operating at the phoneme or character level:

- **Phoneme-level models** exhibit articulatory-like confusions, where phonetically similar sounds (sharing place or manner of articulation) are swapped. This mirrors the error structure seen in dysarthric speech and is consistent with how motor speech cortex encodes articulation.
- **Character-level models** show lexical and word-boundary errors — mistakes that reflect the model's difficulty segmenting continuous neural activity into discrete grapheme sequences without the phonological scaffolding.

This distinction has direct clinical implications. For patients with [Amyotrophic Lateral Sclerosis (ALS)](https://bciintel.com/glossary/als) or high cervical spinal cord injury using intracortical speech BCIs, the failure mode determines how system errors manifest to the user — whether they hear phonetically plausible substitutions or word-level garbling — which shapes both user experience and downstream language model correction strategies.

---

## Why Mamba Didn't Win Here

Selective state-space models like Mamba were developed to handle long-range dependencies in sequence data more efficiently than RNNs, and they have achieved strong results in NLP and genomics. The BCI community has been watching their potential for neural decoding, where temporal dynamics in local field potentials and spike trains span multiple timescales.

The result here — Mamba competitive but not superior — doesn't close the door on state-space models for intracortical decoding. Several considerations are worth flagging:

- **Dataset scale**: Public BCI benchmarks like Brain-to-Text '25 are constrained by the availability of high-quality intracortical recordings from human participants. Mamba-class architectures have often shown advantages at scale; the current dataset size may not favor them.
- **Hybrid architecture**: The authors tested a *hybrid* Mamba decoder, not a pure Mamba stack. Architectural tuning specific to the neural decoding domain remains an open variable.
- **CTC objective**: The choice of training loss constrains the output structure. Whether Mamba's advantages emerge more clearly under sequence-to-sequence or attention-augmented objectives is unresolved.

None of these caveats are speculative additions — they follow directly from the study's own scope. The authors trained under one reproducible protocol precisely to isolate the variables they tested.

---

## Implications for the Speech BCI Pipeline

For the teams at [BrainGate Consortium](https://bciintel.com/companies/braingate), [Precision Neuroscience](https://bciintel.com/companies/precision-neuroscience), and others building intracortical communication BCIs, this paper provides actionable signal: **the GRU-based phoneme decoder remains a defensible architectural choice**, and engineering effort may be better directed at language model integration, electrode count, and signal preprocessing than at swapping the sequence model backbone.

The 21.19% WER from the phoneme-GRU baseline is not yet at the threshold most clinicians would consider natural conversation-quality — general-purpose speech recognition systems operate well below 5% WER on clean audio. The gap reflects both the low signal bandwidth of current intracortical arrays and the challenge of decoding imagined or attempted speech from motor cortex population activity.

Closing that gap will require progress across the full stack: higher-channel-count arrays to capture finer spatial resolution, improved spike sorting and neural preprocessing, more sophisticated language models trained on BCI-specific error distributions, and potentially closed-loop feedback to stabilize neural representations over time.

---

## Key Takeaways

- **GRU outperforms hybrid Mamba** on the Brain-to-Text '25 benchmark under a controlled 2×2 protocol (GRU vs. Mamba; phoneme vs. character targets).
- **Best phoneme-GRU**: 12.62% PER, 21.19% WER. **Best character-GRU** (after LM rescoring): 13.39% CER, 26.28% WER.
- **Mamba is competitive but does not surpass** the recurrent baseline — the advantage of selective state-space models does not transfer cleanly to intracortical speech decoding at current benchmark scale.
- **Failure modes differ by representation**: phoneme models produce articulatory-like confusions; character models produce lexical and word-boundary errors.
- **Clinical takeaway**: Phoneme-level decoding aligns with motor speech cortex encoding and may be more interpretable for patients and clinicians managing system errors.
- This is a **preprint on public benchmark data** — not a live implant study. Independent replication is required.

---

## Frequently Asked Questions

**What is the Brain-to-Text '25 benchmark?**
Brain-to-Text '25 is a public benchmark dataset for evaluating intracortical speech decoding algorithms. It provides standardized neural recordings and transcripts, enabling controlled comparisons across decoder architectures — analogous to how ImageNet standardized computer vision evaluation. It does not involve live implant sessions during the competition evaluation phase.

**What is the difference between phoneme error rate (PER) and word error rate (WER) in BCI speech decoding?**
PER measures the fraction of individual phoneme units decoded incorrectly; WER measures the fraction of complete words in error. WER is the more clinically relevant metric because it reflects how usable the decoded output is as text communication. A decoder can achieve low PER but high WER if errors fall at acoustically critical positions that alter word identity.

**Why might Mamba underperform GRU in neural decoding despite stronger NLP results?**
Several factors may contribute: current intracortical BCI datasets are small by NLP standards, Mamba's efficiency advantages are most apparent at long sequence lengths that may not characterize neural decoding windows, and the CTC training objective used here may not optimally exploit Mamba's selective state-space mechanism. The result is benchmark- and protocol-specific.

**What WER do intracortical speech BCIs need to reach for clinical utility?**
There is no single regulatory threshold, but clinical teams generally target WER below 10–15% for augmentative and alternative communication applications, with the understanding that downstream language model correction and user familiarity with system errors can extend usability above that range. Current best published results on live participants remain above this threshold.

**Does this research apply to ECoG-based or endovascular BCI speech systems?**
The benchmark and study are specific to intracortical (Utah-array style) recordings. [ECoG](https://bciintel.com/glossary/ecog) and endovascular approaches have different spatial resolution, signal bandwidth, and noise characteristics, and architectural findings may not transfer directly. Separate benchmarking on those modalities would be needed to draw equivalent conclusions.