BMMSNet: Bidirectional Mapping and Multilevel Similarity Comparison for EEG-Speech Match-Mismatch Problem

In this report, we present our approach for the task 1 (match-misatch problem) of the Auditory EEG Decoding Challenge at ICASSP 2024. Existing methods for this problem usually only consider the global similarity between EEG and speech stimulus. To overcome the shortcoming, we propose a novel neural network which mainly consists of a Bidirectional Mapping between EEG and speech stimulus and a Multilevel Similarity comparison (BMMSNet). Moreover, we employ a data augmentation-based ensemble learning method to further increase the robustness of our system. Our method achieves a discrimination accuracy of 58.673% on the test data provided by the challenge, ranking third among more than 50 teams. We release the code for our model online.
¹