Under　the　constrained　real-time　condition　of　network,　packet　loss　often　occurs　in　voice　communications　like　voice　over　internet　protocol　(VoIP).　Packet　loss　concealment　(PLC)　techniques　often　use　the　previous　packets　to　recover　the　lost　packet　for　improving　the　quality　of　speech　communication.　In　this　paper,　a　novel　deep　PLC　approach　is　proposed,　which　uses　a　called　Demucs　network　structure,　i.e.,　a　deep　U-Net　architecture　with　a　long-short　time　memory　(LSTM)　network,　to　predict　the　lost　packet　in　the　time　domain.　Firstly,　by　combing　the　convolutions　with　gated　linear　unit　(GLU),　the　encoder　of　network　can　systematically　extract　the　high-level　feature　of　each　speech　frame.　Secondly,　the　LSTM　layers　are　used　to　learn　the　long-term　dependencies　of　speech　frames.　Finally,　the　U-Net　architecture　of　the　network　is　used　to　improve　the　gradient　of　information　flow　by　using　skip　connections,　which　enhances　the　decoder’s　ability　of　reconstructing　the　lost　speech　frames.　Additionally,　the　proposed　architecture　is　optimized　by　utilizing　multiple　loss　functions　in　the　time　and　frequency　domains.　The　experimental　results　show　that　the　proposed　method　has　better　performance　in　perceptual　evaluation　of　speech　quality　(PESQ)　and　short-term　objective　intelligibility　(STOI).　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.　2024.