Nowadays,　due　to　the　application　of　deep　neural　network　(DNNS),　speech　enhancement　(SE)　technology　has　been　significantly　developed.　However,　most　of　current　approaches　need　the　parallel　corpus　that　consists　of　noisy　signals,　corresponding　speech　signals　and　noise　on　the　DNNs　training　stage.　This　means　that　a　large　number　of　realistic　noisy　speech　signals　is　difficult　to　train　the　DNNs.　As　a　result,　the　performance　of　the　DNNs　is　restricted.　In　this　research,　a　new　weakly　supervised　speech　enhancement　approach　is　proposed　to　break　this　restriction,　using　the　cycle-consistent　generative　adversarial　network　(CycleGAN).　There　are　two　stage　for　our　methods.　In　training　stage,　a　forward　generator　is　employed　to　estimate　ideal　time-frequency　(T-F)　mask　and　an　inverse　generator　is　utilized　to　acquire　noisy　speech　magnitude　spectrum　(MS).　Additionally,　two　discriminators　are　used　to　distinguish　the　real　clean　and　noisy　speech　from　generated　speech,　respectively.　In　enhancement　stage,　the　T-F　mask　is　directly　estimated　by　using　the　well-trained　forward　generator　for　speech　enhancement.　Experimental　results　indicate　that　our　strategy　can　not　only　achieve　satisfied　performance　for　non-parallel　data,　but　also　acquire　the　higher　score　in　speech　quality　and　intelligibility　for　the　DNN-based　speech　enhancement　using　parallel　data.　©　2020　IEEE.