With　the　development　of　deep　learning,　automatic　music　transcription　has　witnessed　significant　advancements　in　recent　years.　In　this　work,　we　develop　an　attention-based　model　with　a　fusion　mechanism　to　capture　and　integrate　information　from　time　and　frequency　domains.　Furthermore,　we　adopt　a　separation　strategy　to　avoid　mutual　interference　among　different　branches　in　a　multi-task　model.　Experiments　conducted　on　the　MAESTRO　dataset　demonstrate　that　our　proposed　model　achieves　state-of-the-art　transcription　performance　across　multiple　metrics　among　existing　multi-task　models.　©　2025　SPIE.