Time2Vec Transformer for Robust Gesture Recognition from Low-Density sEMG

OBJECTIVE: To determine whether temporal embeddings can recover the discriminative information lost when reducing sEMG sensing from dense electrode arrays to a sparse two-channel configuration, and under what conditions they can be integrated without degrading spatial features. APPROACH: Using a publicly available dataset of 8 subjects performing 10 dynamic finger gestures, we develop a hybrid Transformer optimized for two-channel sEMG. We identify a failure mode in standard additive integration, where the two branches differ in latent magnitude by a factor of four, so the spatial branch dominates the sum. We therefore integrate Time2Vec embeddings through a normalized additive fusion strategy that layer-normalizes both latent distributions before integration. A two-stage curriculum (augmentation-driven pre-training followed by clean-signal fine-tuning) supports robust feature extraction under data scarcity. MAIN RESULTS: Under leave-one-subject-out cross-validation the proposed model achieves a mean F1-score of 95.9% ± 0.15%, outperforming standard additive, gated, and cross-attention fusion. A gain sweep and a parameter-free normalization variant identify scale alignment as the operative mechanism, as a learnable basis performs worse than fixed sinusoidal encodings under standard addition but exceeds them once normalized. Zero-shot transfer to unseen subjects yields 21.0% ± 2.98%, and classical hand-crafted-feature models collapse comparably, consistent with a sensing constraint rather than a representational one. Supervised calibration with two trials per gesture recovers performance to 96.9% ± 0.52%. Inference runs in 21.5ms on a consumer CPU core. SIGNIFICANCE: Temporal embeddings can compensate for reduced spatial sensing density provided their integration preserves the scale of both representations, enabling rapidly personalizable myoelectric interfaces.