With　the　constructing　of　large　volume　video　dataset　and　the　rapid　development　of　machine　vision　technology,　action　recognition　in　videos　has　become　a　hot　topic　in　many　applications.　The　appearance　and　tempo　of　human　actions　is　variant　in　spatial　and　temporal　space.　It’s　necessary　to　focus　on　the　detailed　descriptions　of　the　fast　and　slow　variations　in　an　action　category.　We　embed　the　spatial–temporal　attention　module　into　the　SlowFast　networks,　to　increase　the　describing　ability　of　fast　and　slow　motion　changes.　The　accuracy　of　the　proposed　method　is　effectively　improved　on　UCF-101　and　HMDB-51　datasets.　Experiments　validate　the　effectiveness　of　the　networks　with　embedded　spatial–temporal　attention　module　for　discriminating　variant　motion　tempos.　The　proposed　method　is　able　to　capture　more　detailed　description　of　action　categories　from　the　slow　and　fast　pathways　and　present　a　more　semantic　recognition　result.　©　2023,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.