← SSI archive · Review rubric

2025 · arXiv · Field expert review · confidence high

A cross-species neural foundation model for end-to-end speech decoding

Yizi Zhang, Linyang He, Chaofei Fan, Tingkai Liu, Han Yu, Trung Le, Jingyuan Li, Scott Linderman, Lea Duncker, Francis R Willett, Nima Mesgarani, Liam Paninski

Introduces a cross-species pretrained transformer encoder enabling state-of-the-art end-to-end neural speech decoding with audio-LLMs, improving accuracy and enabling imagined speech decoding, but latency and real-time deployment remain challenges.

Verdict: full-text draftPriority: highConfidence: highBasis: full text + summaryCoverage: high

Reading guidance

Verdict
full-text draft · priority high · confidence high
Why it matters
Establishes transformer-based cross-species self-supervised pretraining combined with audio-LLM end-to-end decoding as a new paradigm for speech BCIs achieving top decoding accuracy and cross-task generalization, though practical real-time use and robustness require further work.
What to trust
Basis: full text + summary. Coverage: high. 11 evidence records back the review.
What is weak
End-to-end latency and computational requirements; bidirectional attention inhibits online decoding; large pretrained LLMs require substantial resources; need for large labeled and unlabeled datasets; limited real-time applicability currently. Evaluations focused on two human participants (T12, T15) and limited imagined speech vocabulary; unknown performance on larger, more diverse subject sets; small batch sizes for contrastive learning may limit modality alignment effectiveness; inference strategies like nucleus sampling could be further optimized. End-to-end model latency (~0.95s per sentence) and requirement for bidirectional attention constrain real-time deployment; larger LLMs unsuitable for on-device use; dependency on large unlabeled and labeled datasets for pretraining and fine-tuning. Limited to Utah array intracortical neural recordings, tested only on two human participants and associated monkey data; generalization to other populations or hardware untested. Overclaim risk: Low.
Read before
SSI review rubric
Read next
SSI archive

Axes

Task
speech-recognition
Modality
electrophysiological neural spiking activity (Utah arrays)
Hardware
Utah microelectrode arrays with thresholded spikes and spike-band power (SBP) features
Body site
brain
Output
text
Vocabulary
natural language English text
Metrics
Word Error Rate (WER) on Brain-to-Text ’24 and ’25 benchmarks; Phoneme Error Rate (PER) in phoneme decoding; Representational Similarity Analysis scores; processing latency per sentence (approx. 0.95s end-to-end, 0.24s cascaded)
Evaluation mode
Quantitative performance evaluation on held-out datasets (word error rate), ablation studies, representational similarity analysis, and decoding error analyses.
Review confidence
high
Overclaim risk
Low

Expert take

This paper presents a significant step forward in speech brain-computer interfaces by integrating a transformer-based neural encoder pretrained via self-supervised learning on extensive human and monkey neural datasets with an audio-LLM decoder for end-to-end neural-to-text speech decoding. The approach achieves state-of-the-art word error rates on challenging Brain-to-Text ’24 and ’25 benchmarks, substantially improving upon prior RNN-based cascaded and end-to-end models. Importantly, the model successfully decodes both attempted and imagined speech, exhibiting cross-task representational alignment. However, practical deployment is currently limited by slower inference speeds for the end-to-end approach and bidirectional attention's unsuitability for real-time decoding. Further improvements in LLM decoder design, modality alignment, and handling neural signal variability and plasticity are required for long-term, real-world application. Overall, the approach offers a promising foundation for future scalable, integrated neuroprosthetic speech decoding systems.

True value

Establishes transformer-based cross-species self-supervised pretraining combined with audio-LLM end-to-end decoding as a new paradigm for speech BCIs achieving top decoding accuracy and cross-task generalization, though practical real-time use and robustness require further work.

What changed

Canon before

Prior speech BCIs relied on cascaded RNN encoders decoding phonemes combined with n-gram language models. Limited or no large scale pretraining, transformers not widely used, end-to-end speech decoding with LLMs was emerging but still relied on RNNs and lacked large-scale neural data integration.

Delta from canon

Introduces a pretrained transformer-based neural encoder trained cross-species and cross-task with self-supervised masked modeling, integrated end-to-end with large audio-LMs via contrastive alignment to decode neural activity directly to text, improving accuracy and enabling imagined speech decoding and cross-task generalization.

Position in field

Advances state-of-the-art in intracortical speech BCIs by enabling end-to-end transformer and LLM-based decoding with pretrained neural encoders and cross-modal alignment, moving beyond cascaded and RNN-based methods.

Evidence

“ We introduce an end-to-end BraIn-to-Text (BIT) framework that translates neural activity into coherent sentences using a single differentiable neural network with a cross-task, cross-species pretrained neural encoder, supporting both attempted and imagined speech decoding. ”

author_claim · Abstract, Introduction, Methods, Experiments, Discussion · confidence 1.00

“ BIT Cascaded achieves state-of-the-art WER of 6.35% on Brain-to-Text ’24 hold-out, outperforming previous best 7.98%, and BIT End-to-End reduces prior end-to-end WER from 24.69% to 10.22%. ”

metric · Evaluation · confidence 1.00

“ The neural encoder is a transformer pretrained with self-supervised masked modeling on 367 hours of human and monkey Utah array neural data across speech and motor tasks. ”

fact · Methods · confidence 1.00

“ Speech decoding was conducted on Brain-to-Text Benchmark ’24 and ’25 intracortical Utah array datasets with two human participants (T12 and T15) for attempted speech and a smaller imagined speech dataset involving the same individuals. ”

fact · Methods, Evaluation · confidence 1.00

“ Phoneme decoding uses a 41-token vocabulary including phonemes plus blank and silence tokens, with phoneme error rates (PER) correlated with word error rates (WER) after decoding. ”

fact · Appendix · confidence 1.00

“ End-to-end decoding requires about 0.95 seconds per sentence on average, slower than cascaded decoding at 0.24 seconds, limiting real-time applicability; bidirectional attention in the neural encoder is unsuitable for online decoding. ”

limitation · Discussion · confidence 1.00

“ LLMs of larger scale than 1.5B parameters used here cannot run on-device, limiting mobile real-time applications. ”

deployment_claim · Discussion · confidence 0.90

“ This work combines cross-species, cross-task transformer-based self-supervised pretraining with an audio-LLM end-to-end decoder for neural speech decoding, a novel integration beyond prior cascaded, RNN-based, or purely task-specific models. ”

actual_novelty · Full text · confidence 1.00

“ Evaluations were conducted on two particular human participants (T12 and T15) for attempted speech with large vocabularies and on imagined speech with a reduced 50-word vocabulary, limiting generalization scope. ”

validation_scope · Evaluation · confidence 1.00

“ Contrastive learning is employed to align neural and text embeddings in a shared latent space to improve cross-modal alignment and decoding performance. ”

fact · Methods · confidence 1.00

“ This model was pretrained on thresholded spikes and spike-band power (SBP) features from Utah array intracortical recordings. ”

fact · Methods · confidence 1.00

Limits

Technical limits

End-to-end latency and computational requirements; bidirectional attention inhibits online decoding; large pretrained LLMs require substantial resources; need for large labeled and unlabeled datasets; limited real-time applicability currently.

Evaluation limits

Evaluations focused on two human participants (T12, T15) and limited imagined speech vocabulary; unknown performance on larger, more diverse subject sets; small batch sizes for contrastive learning may limit modality alignment effectiveness; inference strategies like nucleus sampling could be further optimized.

Deployment limits

End-to-end model latency (~0.95s per sentence) and requirement for bidirectional attention constrain real-time deployment; larger LLMs unsuitable for on-device use; dependency on large unlabeled and labeled datasets for pretraining and fine-tuning.

Scope limits

Limited to Utah array intracortical neural recordings, tested only on two human participants and associated monkey data; generalization to other populations or hardware untested.