Cross-modal latent alignment enables efficient compression of wearable sEMG and accelerometer signals.

Efficient compression of multimodal biosignals is critical for bandwidth-constrained wearable devices. Existing methods compress surface electromyography (sEMG) and accelerometer signals independently, ignoring their inherent coupling where muscle activation drives limb kinematics. We propose a cross-modal alignment framework that explicitly aligns the latent representations of these modalities during training through frame-wise correspondence and temporal-dynamics consistency. Experiments on two Ninapro datasets demonstrate significant improvements over state-of-the-art methods across compression ratios from 20 to 90%, achieving up to 28% higher correlation coefficients compared to traditional methods and consistent 1-3 dB SNR gains over deep learning baselines. Downstream gesture classification experiments further confirm that our method preserves task-relevant features, maintaining over 80% accuracy at CR = 40% while baselines require CR≥50% for comparable performance. Inference-time latent analysis reveals that alignment training reduces cross-modal drift by 59-64%, validating that the learned consistency persists after the alignment module is removed. Notably, our alignment module serves purely as a training-time regularizer and is completely removed during inference, ensuring that the deployed encoder-decoder system incurs zero additional computational overhead compared to standard single-modality autoencoders. Studies on five backbone architectures confirm that the proposed alignment framework is backbone-agnostic and broadly applicable. Constraint-aware projections onto representative embedded platforms confirm feasibility for real-time wearable deployment.