Speech Decoding

Southern Federal University recorded directional-word EEG from 22 healthy volunteers: 12 native Russian speakers and 10 native Spanish speakers. Both groups performed six directional words, with an additional “next” command in Russian. Recordings used 38 electrodes at 500 Hz, with concurrent articulation-related EMG in six participants. Random Forest accuracy of 78 ± 4% came from per-participant five-fold cross-validation distinguishing rest, overt speech and inner speech—not six-word recognition. Keypress, muscle and task-order effects require consideration in subsequent analyses.

Researchers report SAGE, a switch-aware EEG-guided soft gating framework for target speaker extraction, posted to arXiv as a preprint on August 3, 2026 and not yet peer reviewed. Under in-trial auditory attention switching, neural noise and intrinsic latency can delay or destabilize attention tracking, and conventional methods often cause discontinuities at switching points. SAGE treats in-trial switching as dynamic selection, generating two candidate speech streams with a robust separator and using an EEG-guided switch-aware gating module to produce smooth fusion weights and suppress transition artifacts. It integrates latency-compensated alignment and an uncertainty-driven conservative strategy, outperforms baselines, achieves 8.67 dB SI-SDR and 88.24% STOI, and reduces average switching latency to 2.04 s.

Surgeons in Michigan have implanted Paradromics' Connexus brain-computer interface in a patient with motor neuron disease who had lost most of her ability to speak. Connexus records from individual neurons through 421 microelectrodes seated 1.5 millimeters into the motor cortex; the signals pass to a transceiver in the patient's chest, where AI converts them into text on a screen. Austin, Texas-based Paradromics received FDA approval in the fall of 2025 to begin the clinical trial.

UC Davis researchers trained a transformer-based speech decoder across six people with intracortical BCIs. The pooled model cut relative word error rates by more than 50% on average compared with subject-only models and, after fine-tuning on fewer than 200 sentences from an unseen user, achieved a word error rate below 7%. The bioRxiv preprint has not been peer reviewed.

Aalto University researchers connected the classic N400 evoked response with semantic-vector decoding in a controlled MEG word-priming experiment. The study involved 25 native Finnish speakers who read word groups with different levels of semantic relatedness. Unrelated primes produced larger N400 responses and supplied the most useful training examples for a decoder mapping distributed MEG activity to semantic vectors. The researchers report that semantic information could be decoded from about 100 to 500 ms after stimulus onset. After the N400 peak, however, neural responses no longer mapped reliably to context-invariant semantic vectors. The result suggests that the end of the N400 window may mark a transition from word-specific representation toward broader contextual meaning. This bioRxiv preprint has not been peer reviewed.

A bioRxiv preprint finds that adding a surprisal predictor on top of GPT-2 contextual embeddings contributes almost nothing extra to predicting high-gamma ECoG responses during speech comprehension.

University of Toronto researchers use diffusion latent representations as an intermediate layer for neural speech decoding, with word error rate varying sharply by timestep.

Grenoble University Hospital has registered a clinical study (NCT07698496) assessing the feasibility and effectiveness of a chronic brain-computer interface for speech rehabilitation in people with locked-in syndrome (LIS), where non-invasive communication carries high cognitive load and existing invasive speech BCIs still rely on percutaneous connectors with infection risk. Using an intracranial epidural BCI paired with a speech synthesizer, the SpeechBCI protocol will test two complementary approaches in the same subject — a speech BCI (primary objective, the BCI_PAROLE device) and a cursor BCI (secondary), both running on the WIMAGINE intracranial epidural system. The trial is in the registration stage and has no clinical results yet.

Researchers recorded magnetoencephalography (MEG) from 18 right-handed participants as they imagined sounds and pictures, then compared two decoders: a convolutional neural network (CNN) and a linear logistic regression model. The CNN decoded both tasks above chance and exceeded 70% accuracy for visual imagery. It still decoded significantly when trained only on cortical regions unrelated to the task, suggesting imagined content is spread across partially overlapping networks rather than confined to a single sensory area, an experimental basis for feeding auditory and visual information into BCI decoders together.

A research team has developed Brain2Qwerty, a deep learning model that decodes sentences people type from non-invasive brain recordings. Thirty-five healthy volunteers typed sentences they had just memorized while their magnetoencephalography (MEG, which measures the brain's magnetic fields) or EEG signals were recorded. The average character error rate was 29% with MEG and 65% with EEG; for the best-performing participants it fell to as low as 18%, and the model fully decoded a range of sentences outside its training set. The researchers say the results narrow the gap between non-invasive and implanted methods.

A study published in Computers in biology and medicine proposes dynamic wavelet-basis selection to improve non-invasive EEG imagined-speech classification. For each EEG epoch, the method minimizes wavelet entropy to select an informative basis and then injects Gaussian noise into the corresponding coefficients. A convolutional neural network with channel-wise excitation classifies the augmented signals. The dataset contains 32 channels, 8 stimuli, and recordings from 10 participants. The words-vowels combination reached a highest classification accuracy of 98% with a Cohen's kappa of 0.95, although performance was lower for the full class set. The researchers report that the method outperformed conventional augmentation strategies and static wavelet approaches, offering an adaptive way to address noise and non-stationarity in EEG decoding.

A team led by the UMC Utrecht Brain Center in the Netherlands analyzed 8 participants with high-density electrocorticography (ECoG) grids of 64 to 128 electrodes and found that a well-positioned 32-electrode grid classified 12 Dutch words with 76 ± 16% accuracy, comparable to 75 ± 17% for the larger grids. The best placement was vertical and centered on the central sulcus. Speech brain-computer interfaces aim to restore communication to people paralyzed and unable to speak, but larger grids carry higher surgical risk. The participants were not paralyzed and testing was limited to a fixed word set.

A Neurosurgical Focus paper tested Precision Neuroscience's 1,024-channel Layer 7 micro-ECoG array in four awake-craniotomy patients. Four-word speech classification reached 77.5% accuracy and four-direction cursor classification reached 78% to 84%, with no device-related adverse events reported during the procedures. The small, short intraoperative study was a feasibility test rather than a pivotal motor-restoration trial.