FETTUCCINE: Fast and efficient brain-to-text decoding on mobile devices

Jonathan D. McCart, Pranav I. Deevi, Mehdi Azabou, Nanda H. Krishna, Caleb A. McKinney, Nicholas S. Card, Sergey D. Stavisky, Chethan PandarinatharXiv 2026 · arXiv

TL;DR

FETTUCCINE is a new end-to-end framework that brings high-performance brain-to-text decoding to off-the-shelf mobile devices. By adapting pre-trained Automatic Speech Recognition (ASR) models, the researchers achieved word error rates (WER) as low as 4.39% while reducing memory requirements from hundreds of gigabytes to less than 3GB, enabling the first-ever demonstration of real-time neural speech decoding on a smartphone.

What Problem Does This Paper Address?

Speech neuroprostheses—devices that translate brain activity into text—have shown incredible promise for restoring communication to individuals with paralysis or speech impairments. However, current state-of-the-art systems are "cascaded": they use a neural decoder to predict phonemes, which are then processed by a massive stack of language models (n-grams and LLMs) to produce final text.

This architecture creates two major barriers to real-world use:

  1. Infrastructure: These systems often require dedicated workstations with over 300 GB of RAM or persistent cloud access, tethering users to research labs.
  2. Latency: The language modeling step (specifically beam search over n-gram graphs) suffers from highly variable inference times. A single sentence might take 50ms or over 10 seconds to decode, making natural, fluid conversation impossible.

Methods

The authors introduce FETTUCCINE (Fast and Efficient Thought-to-Text Unified Cross-modal Communication Interfacing via Neural Embeddings). Instead of building a massive cascaded stack, they leverage the efficiency of existing ASR models like OpenAI’s Whisper and NVIDIA’s Parakeet.

Figure 1: Architecture and training strategy. (Top) Our framework modifies open-source ASR models using a Neural Adapter network and LoRA updates targeted at the ASR encoder layers. Train- ing occurs in two stages: a warm-up isolating the Neural Adapter, followed by joint training of the adapter and LoRA updates. The remaining ASR backbone remains frozen. (Bottom) A cascaded ap- proach for comparison, featuring a non-end-to-end pipeline with a distinct GRU phoneme decoder.

Key technical innovations include:

  • Neural Adapter: They replaced the standard audio convolution front-end of ASR models with a custom "Neural Adapter" (a temporal U-Net variant) that maps high-dimensional neural signals directly into the ASR's embedding space.
  • Two-Stage Training: The model undergoes a "warm-up" phase for the adapter, followed by joint training with the ASR backbone using LoRA (Low-Rank Adaptation) to maintain efficiency.
  • Neural Envelope Augmentations: Since brain data lacks the clear word boundaries found in audio, the team developed a way to estimate "neural envelopes"—proxies for word timings—to apply word-level data augmentations like warping and cropping, significantly improving robustness.

Datasets

The framework was evaluated on two major benchmarks:

  • Brain-to-text 2025 Challenge (Participant T15): Recordings from four 64-channel microelectrode arrays in the speech motor cortex, spanning 20 months and ~11,000 sentences.
  • Brain-to-text 2024 Challenge (Participant T12): A similar dataset from a second participant with implants in area 44 and 6v, spanning 4 months and ~10,000 sentences.

Key Insights and Ideas

  • ASR Models are Robust Cross-Modal Learners: A surprising "aha" moment is that ASR models, originally trained on millions of hours of audio, can be effectively "re-wired" to interpret neural spikes with minimal fine-tuning.
  • Efficiency is the Key to Portability: By moving to an end-to-end architecture, the authors eliminated the memory-hungry n-gram graphs. This allowed them to fit the entire model into the Unified Memory Architecture of an iPhone.
  • Grounding via Saliency: Using gradient-based saliency maps, the authors proved the models weren't just "hallucinating" likely sentences based on language priors. Instead, the models were attending to specific electrodes in canonical speech motor regions, aligning perfectly with physiological expectations.
Figure 5: Saliency maps showing the model correctly localizing neural features in the motor cortex.

Key Results

  • State-of-the-Art Accuracy: FETTUCCINE achieved a 4.39% WER on the T15 dataset, outperforming previous end-to-end methods.
  • Mobile Breakthrough: The models successfully ran on an iPhone 15 Pro Max, achieving throughput up to 370x real-time.
  • Massive Memory Reduction: While traditional cascaded models require 320 GB of RAM, FETTUCCINE variants (like Whisper Tiny) require as little as 209 MB of peak memory.
  • Consistent Latency: Unlike cascaded models with "long-tail" latencies, FETTUCCINE provides fast, predictable inference, which is critical for user safety and natural interaction.
Figure 3: Brain-to-text 2025 val WER vs. Peak Memory. Shaded: mobile- ready models (usable and portable).

Limitations and Future Work

While FETTUCCINE is a leap forward for portability, the authors note a few areas for improvement:

  • Absolute Accuracy: The very best cascaded systems (using massive server-side LLMs) still achieve slightly lower WERs (around 2-3%). For high-stakes medical or legal communication, users might still prefer the bulkier systems.
  • Calibration Time: The models currently require roughly 5–10 sessions of data to reach usable performance levels. Reducing this "warm-up" period for new users is a priority.
  • Acausality: Current variants process the entire sentence at once. Future work will explore "streaming" or causal backbones to allow for word-by-word decoding as the user thinks them.

Summary updated

Abstract

Recent advances in speech neuroprostheses have enabled new communication avenues for people who have lost the ability to speak due to neurological illness or injury. These systems can decode neural activity during attempted speech into text by relying on language priors to achieve high decoding accuracy. This performance currently comes at the cost of using a compute-intensive language modeling stack, presenting a major bottleneck for real-world use. Further, current language models incur highly variable and disruptive inference latencies, limiting communication throughput and preventing users from reliably participating in natural conversation. We introduce FETTUCCINE, an end-to-end brain-to-text framework that leverages the speed and portability of automatic speech recognition (ASR) models to address the aforementioned practical challenges, while achieving usable (<10%) word error rates (WER), efficient adaptation to neural non-stationarities, and reliable performance on mobile devices. When evaluated using publicly available data from an intracortical speech neuroprosthesis user, FETTUCCINE outperforms previous end-to-end methods, achieving as low as 4.39% WER. Most notably, our models successfully run on a commercially-available mobile device with throughput reaching up to 370x real-time, providing the first demonstration of brain-to-text decoding on a mobile device. We also show that our models can be successfully finetuned to extend usable performance to future days. Beyond performance, our approach enables localization of the neural features most relevant for decoding via gradient-based saliency maps, which we show align with well-established physiological priors across different ASR model families. Taken together, our results show that FETTUCCINE overcomes core infrastructural and computational barriers, yielding a new class of portable brain-to-text communication.
BibTeX
@misc{mccart2026fettuccine,
  author = {Jonathan D. McCart and Pranav I. Deevi and Mehdi Azabou and Nanda H. Krishna and Caleb A. McKinney and Nicholas S. Card and Sergey D. Stavisky and Chethan Pandarinath},
  title = {{FETTUCCINE: Fast and efficient brain-to-text decoding on mobile devices}},
  year = {2026},
  eprint = {2611.18561},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18561},
}
Figure