An optimized heuristic domain adaptive structure for EMG-based cross-subject silent speech recognition

EMG-based silent speech recognition has achieved favorable performance recently, providing the possibility of assisted therapy for speech disorders and communication in specialized settings. Aiming at handling the dramatic degradation of classification accuracy caused by biological variations and articulation bias across speakers, we propose an optimized heuristic domain adaptive architecture extracting global features, speaker-related domain-specific features, and speakerindependent domain-invariant features from myoelectric signals, respectively. Through separate alignment and optimization in parallel, the effectiveness of the extracted features and the classification performance have been significantly enhanced. Experimental results show that our method achieves a classification accuracy of 94.40%, exceeding the state-of-the-art model by almost 10 percent, in a scenario with 60 subjects for training and 10 new subjects for validation.