By　taking　into　account　temporal　correlation　of　speech　feature.　In　this　paper,　a　novel　structure　of　convolutional　Auto　Encoder　(CAE)　was　proposed.　In　this　structure,　the　historical　output　of　the　CAE　was　fed　into　a　CAE　stack　recurrently.　We　name　this　structure　as　Recurrent　Stack　Convolutional　Auto　Encoder　(RS-CAE).　In　the　training　stage,　the　training　feature　maps　of　the　RS-CAE　comprise　of　log　power　spectrum　(LPS)　of　noisy　speech　and　an　additional　feature　map　derived　from　the　LPS　of　the　enhanced　speech　in　the　history.　In　this　way,　the　temporal　correlation　is　incorporated　as　much　as　possible　in　the　RS-CAE.　The　training　target　is　a　concatenated　vector　of　auto-regressive　(AR)　model　parameters　of　speech　and　noise.　At　online　stage,　the　LPS　of　noisy　speech　and　the　LPS　of　the　enhanced　speech　from　the　history　make　up　input　feature　maps　together.　The　outputs　of　the　RS-CAE　are　the　AR　model　parameters　of　speech　and　noise,　which　are　used　to　construct　the　AR-Wiener　filter.　Because　the　estimated　AR　model　parameters　are　not　completely　accurate　and　some　harmonics　may　be　lost　in　the　enhanced　speech,　the　codebook-based　harmonic　recovery　technique　was　proposed　to　reconstruct　harmonic　structure　of　the　enhanced　speech.　The　test　results　confirmed　that　the　proposed　method　achieved　better　performance　compared　with　some　existing　approaches.