Brain-to-text decoding runs on a phone, in offline tests

A preprint adapts off-the-shelf speech recognition models to turn recorded brain activity into text on an iPhone. It uses a small fraction of the memory that current research systems need. It has not yet been tried with a live user.

6 min readIIBenchmarkedPreprint, not peer reviewed
Save
A tall stack of wooden crates bound with string sits beside a tiny, open matchbox holding folded blue paper.AI illustration
Illustration: AI generated

What the study shows

Speech neuroprostheses are brain implants that read the activity of speech areas while a person tries to talk. Software then turns that activity into text. Most current systems work in stages. One model guesses speech sounds. Then large language models, which predict likely word sequences, turn those sounds into sentences. The language stage alone can need about 320 GB of computer memory.

The authors take a different route. They start with open speech recognition models, Whisper from OpenAI and Parakeet from NVIDIA. These normally turn audio into text. They swap the audio input layer for a small trained adapter that feeds brain signals in instead. They also lightly tune the model with LoRA, a method that trains a small set of extra weights rather than the whole model. Text comes out in one step, with no separate language stack. They call the framework FETTUCCINE.

They tested it on the public Brain-to-Text 2025 benchmark. This holds recordings from one participant with ALS, known as T15, copying cued sentences. The best variant got 4.39% of words wrong on the public test split. The usual staged system scored 6.67%. The best earlier one-step system in their comparison scored 8.70%. Their own models needed about 1 to 5 GB of memory on a server GPU, against about 320 GB for the staged baseline.

The authors then ran six variants on an iPhone 15 Pro Max, on the benchmark's validation set. The smallest model used about 209 MB of memory. It got 8.4% of words wrong and took under a twentieth of a second per sentence. The authors say this is the first brain-to-text decoding on a mobile device. That claim is based on their own knowledge of the field.

The models do not transfer to new days on their own. Brain signals drift over time. On five later sessions, error without adjustment ranged from 13% to 79% for one variant. Retraining on that day's data brought it down to between 3% and 13%. On average, about 15 minutes of new recordings, around 50 sentences, gave reasonable results.

Results were weaker for a second participant, T12, from the 2024 benchmark. There, the staged systems still did better, and the authors call these results preliminary. The paper is a preprint and has not been peer reviewed.

Our reading

In our reading, the main result is about cost and portability, not accuracy. The accuracy is good for a one-step system. It is not the best on the benchmark: the authors note the top five competition entries reach 2% error or lower. What changes is the hardware bill. A decoder that fits in a few hundred megabytes removes the need for a workstation or a cloud link.

The steadiness of response time may matter as much as speed. The staged baseline averaged about half a second per sentence, but some sentences took more than 10 seconds. That happens when the word search expands on noisy input. The new models stayed close to their average. For someone using an implant to ask for help, a reliably quick answer is worth more than a fast average with long stalls.

A concrete example: a person with ALS at home could carry the decoder on a phone, with no server and no internet. Their brain data would stay on the device. That is the authors' stated aim. This study shows the computing part is feasible on recorded data. It does not yet show the whole loop working with a person. On our four-level scale, we rate it II, Benchmarked: there are public benchmark results, but no released code or weights.

Free membership

Read the rest of this story

Membership is free. You get every article in full and the Issue every Sunday.

We send a sign-in link, no password needed. Already a member? Sign in

Readiness assessment

II
Benchmarked
LevelStatusEvidence
I  Lab onlyMetA working method with results on recorded data from two participants, and a demonstration on an iPhone.
II  BenchmarkedMetResults on the public Brain-to-Text 2025 and 2024 benchmarks, compared with published baselines (Table 1). Caveat: some baseline numbers are copied from other papers.
III  Open sourceNot yetThe checklist says the full training codebase cannot be released because of an IP review. No weights are released. A project page is named (fettuccine-brain2text.github.io), but the paper does not say it hosts code or weights. The underlying ASR models and datasets are public.
IV  In productionNot yetNo documented use by a named organisation or in closed-loop clinical sessions. The authors list this as future work.

What would move it up
Release of the training code and trained weights would move it to level 3. Level 4 would need documented closed-loop use with participants in a clinical research setting, with streaming (causal) models, and accuracy matching offline results.

The study

Jonathan D. McCart, Pranav I. Deevi, Mehdi Azabou et al.. FETTUCCINE: Fast and efficient brain-to-text decoding on mobile devices. arXiv, 13 November 2026. Read the paper

Preprint, not peer reviewed · Licence CC BY NC ND 4.0

Notes

  1. McCart, Deevi et al., FETTUCCINE: Fast and efficient brain-to-text decoding on mobile devices. arXiv preprint 2611.18561, November 2026. Not peer reviewed. Listed for NeurIPS 2026. Source
  2. Project page named in the paper. The paper does not say it hosts code or weights; the training code is withheld pending IP review. Source
  3. Exact figures, Brain-to-Text 2025 test (T15, 1,450 sentences): Whisper Small 4.39% public, 5.05% private; cascaded GRU + 5-gram + OPT-6.7B 6.67% and 7.00%, about 320,000 MB RAM and 18,100 MB GPU memory, latency 494.8 ± 2457.7 ms; BrainWhisperer 8.70%; BIT 11.06%.
  4. Exact phone figures, iPhone 15 Pro Max, 2025 validation set: Whisper Tiny 8.4% WER, 48.4 ms, about 373 times real time, 209 MB peak; Whisper Large v3 Turbo 5.2% WER, 2,959.8 ms, 676 MB with encoder split in two.
  5. Brain-to-text '25 benchmark on Kaggle. Source
  6. Readiness: II, Benchmarked. Public benchmark results; no released code or weights; no documented live or clinical use.

Edited and signed by

Stefan PerdalEditor and publisher86 stories edited

Stefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.

How this was made

An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.

More on Efficiency