In　sleep　environment,　respiratory　signals　are　inevitably　affected　by　ambient　noise.　In　order　to　improve　the　noise　estimation　ability　of　noise　spectrum　estimation　algorithm,　an　improved　IMCRA　non-stationary　noise　estimation　algorithm　is　proposed.　The　algorithm　uses　the　3　spectral　smoothing　method　to　estimate　the　probability　of　respiratory　signals　and　control　the　length　of　segments.　In　the　case　of　noise　overestimation　or　underestimation,　a　safety　mechanism　is　introduced　to　limit　or　compensate　for　the　noise　estimate.　Experimental　results　show　that　compared　with　IMCRA　algorithm,　the　improved　algorithm　improves　the　noise　estimation　ability.　At　the　same　time,　the　intelligibility　and　sharpness　of　speech　signals　are　improved,　which　can　be　applied　to　de-noising　sleep　signals　better.　©　2018　IEEE.