Effectiveness of single-model approaches for silent speech decoding under heterogeneous EEG/EMG electrode configurations and transfer to patients

Determine the effectiveness of single-model deep neural network architectures that handle heterogeneous EEG and electromyography (EMG) electrode configurations for silent speech decoding tasks, and ascertain the efficacy of knowledge transfer from models trained on healthy individuals to patients with neurodegenerative diseases, when using non-invasive EEG/EMG signals.

Background

The paper discusses prior work on generalizable decoders trained across diverse EEG electrode configurations, which showed utility on tasks like motion or imagined motion detection. However, silent speech decoding is a more complex downstream task and involves both EEG and EMG signals with varied electrode placements across devices and subjects.

Motivated by this gap, the authors introduce and evaluate four tokenizers (global average pooling, electrode-specific, subject-specific, and an on-the-fly kernel) within a unified architecture, training on large-scale heterogeneous datasets from eight healthy participants and one patient. The stated uncertainty concerns whether such single-model approaches are effective for silent speech decoding and whether knowledge learned from healthy participants transfers to patients.

References

While these studies demonstrated the utility of handling heterogeneous electrode configurations within a single model to increase available data for tasks such as motion/imagined motion detection, the effectiveness for relatively complex tasks such as silent speech decoding remains unclear, as does the efficacy of knowledge transfer between healthy individuals and patients.

— A Silent Speech Decoding System from EEG and EMG with Heterogenous Electrode Configurations (2506.13835 - Inoue et al., 16 Jun 2025) in Section 1 (Introduction)