Wavelet Scattering Features for Robust Subject-Independent EMG-Based Speech Recognition: A Comparative Analysis

Electromyography (EMG)-based speech recognition provides a non-acoustic alternative in high-noise environments and serves as a prospective foundation for assistive technologies for users with speech impairments. While prior work has largely focused on model architecture, the impact of EMG feature representations on the robustness and subject-independent generalization of deep learning models remains underexplored. This study aims to investigate how feature representation influences subject-independent EMG speech recognition by comparing two feature extraction pipelines-a conventional Mel-Frequency Spectral Coefficients (MFSC) pipeline and a Wavelet Scattering Transform (WST)-based pipeline-across four temporal back ends: gated recurrent unit (GRU), Transformer, ReMamba, and BiMamba. Recordings from 100 participants during voiced (vocalized) speech reading a closed set of fixed sentences were used, of which 20 participants' recordings were excluded from model development and used exclusively for subject-independent testing to verify generalization. The speaker dependence of both representations was additionally examined with a participant-level variance analysis. The WST-based pipeline achieved higher sentence-level accuracy than the MFSC pipeline with every backend, with the best result reaching 85.82% on unseen participants. For the Transformer backend, accuracy increased from 78.59% with MFSC to 85.82% with the WST pipeline. Speaker identity explained 12% of the feature variance in the WST pipeline compared with 69% for MFSC. When the filtering pipeline used for MFSC was applied before scattering extraction, accuracy decreased by 2.16 percentage points, indicating that this preprocessing may suppress EMG information useful to the model. Collectively, the findings identify the WST-based pipeline as a less speaker-dependent EMG feature representation that improves cross-subject performance across different sequence models and may provide a basis for intelligent EMG speech-recognition systems.