Silent Speech Interfaces (SSIs) can be extremely useful for individuals suffering from auditory or vocal impairments. By leveraging surface electromyography (sEMG) to capture the localized bio-electrical excitation of facial and articulatory muscles, these interfaces establish an alternative, non-acoustic communication channel that is completely robust to environmental noise and viable for speech-impaired populations. However, the deployment of deep learning models on sEMG data is severely constrained by low signal-to-noise ratios, cross-subject physiological variability, and a scarcity of labelled EMG datasets relative to massive acoustic speech corpora. To systematically address these bottlenecks, this research explores two distinct engineering tracks. Track A investigates a closed-vocabulary cross-subject Mandarin sentence classification task using the contemporary AVE Speech dataset.