Neural decoder
Also: decoding algorithm, BCI decoder, brain decoder
By neuraspeak editorial · Updated 2026-10-07 · 1 min read
A neural decoder is the software model in a brain-computer interface that translates recorded brain activity into an intended output, such as a cursor velocity, a letter, a phoneme, a word, or sound, usually by learning the mapping from examples collected from the user.
How is a neural decoder trained?
The user is shown prompts, for example a sentence to attempt to say, while their neural activity is recorded. Because the system knows what the user was trying to produce, each stretch of brain data gets a label. A model is then fit to predict the label from the activity. Early BCIs often used linear models; current speech systems typically use recurrent neural networks. Willett et al. (Nature 2023) trained a recurrent network that output, every 80 milliseconds, the probability of each phoneme being spoken. Card et al. (NEJM 2024) reported that 30 minutes of calibration data on the first day was enough for 99.6% accuracy on a 50-word vocabulary.
Why do speech decoders use a language model?
Decoded phoneme probabilities are noisy. A language model, of the same general kind used in speech recognition and text prediction, scores which word sequences are plausible in English and picks the most likely sentence consistent with the neural evidence. Moses et al. (NEJM 2021) combined word classifiers with a language model to decode sentences, and later systems pair phoneme decoders with large-vocabulary language models. This step greatly reduces errors, but it also means some output reflects language statistics, not neural data alone.
Does a decoder stay accurate over time?
Not automatically. Neural recordings change from day to day, a problem called neural signal drift, so decoders are periodically recalibrated or continuously updated. Willett et al. used day-specific input layers and rolling adaptation; Card et al. combined data across days and offered on-demand recalibration taking about 7.5 minutes.
Why it matters for speech BCI
In a speech neuroprosthesis, the decoder decides almost everything the user experiences: speed, word error rate, vocabulary size, and latency. Advances in deep learning and language modeling are a large part of why speech BCI performance rose from about 15 words per minute in 2021 to over 60 words per minute in 2023.
Questions
Is a neural decoder the same as AI?+
A modern decoder is a machine-learning model, so it is a form of AI. It is trained on one person's brain data for a specific task, unlike a general-purpose chatbot.
Does the decoder work for anyone right away?+
No. Current decoders are trained on each user's own recordings, because electrode placement and brain activity patterns differ between people.
Related: Phoneme decodingNeural signal driftSpeech neuroprosthesisSpike sortingBrain-computer interface (BCI)
Sources
- Willett et al., A high-performance speech neuroprosthesis, Nature 620, 1031-1036 · 2023
- Card et al., An Accurate and Rapidly Calibrating Speech Neuroprosthesis, New England Journal of Medicine 391, 609-618 · 2024
- Moses et al., Neuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria, New England Journal of Medicine · 2021
- Metzger et al., A high-performance neuroprosthesis for speech decoding and avatar control, Nature 620, 1037-1046 · 2023
Facts on this page are drawn from the sources listed and were accurate to the best of our knowledge on the update date shown. Company and product names are trademarks of their owners and are used for identification only. Nothing here is medical, investment or legal advice. Companies and individuals can request a correction via the corrections page.