Emotion　Recognition　in　Conversation　(ERC)　has　emerged　as　a　pivotal　topic　in　the　realm　of　human-computer　interaction,　drawing　escalating　attention.　Despite　previous　research　achieving　certain　accomplishments,　most　approaches　treat　each　modality　equally,　failing　to　differentiate　the　emotional　information　across　different　modalities　and　thus　struggling　to　harness　the　complementary　and　associative　information　within　multimodal　data.　To　address　this　issue,　this　paper　propose　a　Cross-Modal　Fusion　Network　with　Gated　Units　(CFN-GU).　CFN-GU　comprises　two　main　components:　the　Single-Modal　Transformer　and　the　Learnable　Fusion　Strategy　With　Gate　(LG-Fusion).　The　Single-Modal　Transformer　is　employed　to　model　contextual　information　for　each　unimodal　feature,　extracting　rich　contextual　emotional　cues.　Subsequently,　LG-Fusion　autonomously　learns　the　weights　of　each　feature　information　for　every　modality,　thus　comprehensively　understanding　the　contributions　of　different　modalities　to　emotional　information.　Finally,　the　information　from　the　three　modalities　is　fused　based　on　these　learned　weights.　CFN-GU　achieves　an　F1　score　of　64.3%　on　IEMOCAP,　effectively　improving　ERC　performance　and　outperforming　all　benchmark　baselines.　　©　2024　IEEE.