With　the　increase　of　microphone　array　applications,　multi-channel　speech　coding　has　become　an　important　technology.　In　this　paper,　using　the　powerful　learning　ability　of　deep　neural　network,　a　new　four-channel　speech　coding　method　based　on　recurrent　neural　network　(RNN)　is　proposed.　Only　two　signals　of　four　channels　are　encoded　and　decoded　separately　using　a　standard　codec.　These　two　decoded　signals　are　processed　to　obtain　their　residual　signals　by　using　the　well-known　linear　prediction　analysis.　The　RNN　is　used　to　learn　the　differences　lied　in　the　residual　signals　in　the　training　stage　and　predict　the　residual　signals　of　other　two　channels　in　the　testing　stage.　Finally,　speech　signals　of　the　other　two　channels　can　be　recovered　using　the　synthesis　process.　Through　objective　measurements,　the　performance　of　the　proposed　method　was　assessed　and　the　experimental　results　indicated　that　the　proposed　method　could　obtain　a　higher　efficiency　for　four-channel　speech　coding.　©　2021　IEEE.