Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study benchmarks EEG foundation models against established convolutional networks for speech decoding, utilizing two distinct corpora: UGR-MINDVOICE and BCI Competition 2020 Track 3. Despite the large-scale pretraining of models like LaBraM and EEGMamba, the results reveal no consistent performance advantage over a simpler 16K-parameter CNN. These findings suggest that current EEG pretraining methods are inadequate for the complexities of speech production, highlighting the need for models specifically designed for this task.
Current EEG foundation models fail to outperform basic CNNs in speech decoding, signaling a critical gap in their applicability to one of the most challenging BCI tasks.
EEG foundation models pretrained on thousands of hours have shown large gains over task-specific networks for motor imagery, seizure detection, sleep staging, and emotion recognition, but their transfer to speech decoding-arguably the most demanding non-invasive BCI application-remains untested. We present the first systematic benchmark of EEG foundation models against strong convolutional baselines for speech decoding, using two corpora: UGR-MINDVOICE (overt and covert Iberian Spanish) and BCI Competition 2020 Track 3 (imagined speech). We compare two foundation models (LaBraM, EEGMamba) against three established baselines (EEGNet, ShallowFBCSPNet, EEGConformer) under a unified preprocessing and fine-tuning protocol. Large-scale EEG pretraining yields no consistent advantage over a 16K-parameter CNN on speech tasks, indicating that current general-purpose EEG pretraining does not yet transfer to speech production and motivating speech-specific foundation models.