A　fast　convergence　speech　enhancement　method　is　proposed　in　this　paper.　The　noise　estimation　acceleration　technique　is　applied　to　the　conventional　statistical　model　based　algorithm　to　shorten　the　convergence　time　after　the　sudden　change　of　noise　intensity.　First,　the　burst　detection　of　power　spectrum　is　performed　on　the　noisy　spectrum.　Next,　the　loglikelihood　ratio　(LLR)　based　VAD　is　used　in　the　period　when　the　noise　power　is　stationary,　and　the　spectral　entropy　based　VAD　is　implemented　in　the　hang-over　frames　after　the　burst　of　noisy　spectrum.　Then　a　flag　is　set　to　control　the　update　of　noise　estimation.　Finally,　an　attenuated　version　of　the　noisy　spectrum　will　be　used　directly　as　noise　estimation　if　the　update　flag　is　set　to　one.　The　performance　of　the　proposed　method　is　evaluated　under　ITU-T　G.160.　In　comparison　with　the　conventional　method,　the　convergence　time　is　reduced　evidently,　while　the　abilities　of　noise　reduction　and　SNR　improvement　are　preserved,　and　the　impact　on　the　objective　speech　quality　is　constrained　to　a　low　level.