STCCA: Spatial-Temporal Coupled Cross-Attention Through Hierarchical Network for EEG-Based Speech Recognition

Liang Dong ¹, Hengyi Shao ¹, Lin Zhang ^1,2

⁰School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876, China.

Sensors (basel, Switzerland) +

|

November 13, 2025

View abstract on PubMed

Summary

This study introduces a novel spatial-temporal coupled cross-attention (STCCA) network for Electroencephalogram (EEG) based speech recognition. STCCA enhances feature fusion, improving accuracy in brain-computer interface applications.

Area Of Science

Neuroscience
Computer Science
Biomedical Engineering

Background

Electroencephalogram (EEG) based speech recognition shows promise for communication and rehabilitation.
Current methods often fail to fully capture complex spatial-temporal feature interactions.

Purpose Of The Study

To propose a novel hierarchical network, STCCA, for improved EEG speech recognition.
To address limitations in existing methods by exploring intricate cross-relationships between spatial and temporal features.

Main Methods

Developed a spatial-temporal coupled cross-attention (STCCA) mechanism within a hierarchical network.
Utilized Convolutional Neural Networks (CNNs) for local feature extraction (LFEM).
Employed a dual-directional attention mechanism for coupled cross-attention (CCA) fusion.
Incorporated multi-head self-attention layers for global feature extraction (GFEM).

Main Results

STCCA achieved accuracies of 45.45%, 25.91%, and 29.07% on three EEG speech datasets.
Demonstrated performance improvements of 1.95%, 3.98%, and 1.98% over existing models.
The coupled cross-attention module effectively modeled deep interactions between spatial and temporal features.

Conclusions

The proposed STCCA network offers a significant advancement in EEG-based speech recognition.
The hierarchical approach with coupled cross-attention effectively captures complex neural signal dynamics.
This method holds potential for enhanced brain-computer interfaces in communication and rehabilitation.

Keywords:

EEG-based speech recognition cross-attention fusion deep learning multi-head self-attention