Efficient sEMG decoding with hybrid, multi-timescale neural networks

Surface electromyography (sEMG) offers a non-invasive window into neuromuscular activity and is increasingly used as a reliable control signal for prosthetics, rehabilitation devices, and wearable health monitoring.However, decoding sEMG for such applications is computationally expensive, which remains a major barrier to deployment on embedded and wearable systems.To overcome this limitation, we introduce novel multi-timescale architectures for large-scale, continuous sEMG decoding.We first show that structured linear state-space models (SSMs), originally developed for long-sequence modeling, provide an efficient baseline for sEMG decoding on the emg2qwerty benchmark.These surpass state-of-the-art, transformer-based performance while requiring up to three orders of magnitude fewer operations.Building on this result, we introduce a hybrid multi-timescale architecture for the same large-scale sEMG decoding task, integrating linear SSMs, convolutional and event-driven networks within a structured two-stage framework.Our results show that the proposed homogeneous SSMs are optimal at high computational budgets (< 100 M MACs/s), whereas the introduced two-stage models substantially improve accuracy–efficiency scaling in the computationally constrained regimes relevant for edge deployment.In this regime, our two-stage models match or exceed prior state-of-the-art accuracy with up to 5000× lower computational cost.Finally, to demonstrate real-time operation in both distributed and end-to-end edge implementations, we deploy our best models on low-power STM32 microcontrollers, validating analytical complexity estimates against measured latencies.Overall, these results demonstrate that task-aligned, computationally efficient architectures enable real-time sEMG decoding on edge devices for naturalistic, dexterous motor control, paving the way towards effective prosthetics and rehabilitation devices, and wearable assistive and augmentative technologies.