Inter-subject　continuous　motion　estimation　using　surface　electromyography　(sEMG)　is　essential　in　rehabilitation　medicine,　as　it　can　substantially　reduce　the　costs　associated　with　personalized　adjustments.　In　this　study,　we　introduce　a　Multi‑Scale　Spatio‑Temporal　(MSST)　model　demonstrating　excellent　feature‑extraction　capabilities.　To　enhance　inter‑subject　accuracy　on　small　datasets,　we　combine　the　gradient　harmonizing　mechanism　(GHM)　with　variational　mode　decomposition　(VMD).　To　rigorously　evaluate　the　model＇s　generalizability,　we　assess　it　under　multiple　conditions:　intra‑subject　and　inter‑speed;　inter‑subject　and　intra‑speed;　and　inter‑subject　and　inter‑speed　scenarios.　In　every　scenario,　MSST　gets　top　performance　and　outperforms　state‑of‑the‑art　models　such　as　BioCNN‑T　and　Conv‑BiLSTM.　We　further　validate　our　approach　using　the　public　Ninapro　DB2　dataset.　The　model　again　yields　superior　performance　across　both　intra‑subject　and　inter‑subject　scenarios.　In　summary,　our　work　proposes　a　robust　framework　for　inter‑subject　motion　estimation　from　sEMG　signals.　©　2025