For　sound　event　detection　under　various　background　noises　at　low　SNR,　this　paper　proposes　a　method　that　mixes　the　background　noises　with　sound　events　into　noisy　samples　to　train　classifiers.　In　the　pre-processing　stage,　we　use　a　voting　method　based　on　2th　to　6th　intrinsic　mode　functions　(IMFs)　that　generated　from　empirical　mode　decomposition　(EMD),　to　detect　the　endpoint　of　sound　events　and　estimate　the　SNR.　Then　subband　power　distribution　(SPD)　is　used　to　extract　features　from　audio　data.　Finally,　we　mix　the　background　noise　and　all　the　sound　event　samples　in　the　sound　event　database　according　to　the　estimated　SNR,　and　then　extract　the　noisy　samples　features　to　train　multi-random　forests　(M-RF)　for　the　detection　of　the　sound　events　in　low　SNR　environment.　The　experiment　proves　that　the　proposed　method　has　the　ability　to　recognize　sound　events　in　various　acoustic　scenes　at　low　SNR,　and　can　remain　an　average　accuracy　rate　of　67.1%　at-5dB.　©　2018,　Chinese　Institute　of　Electronics.　All　right　reserved.