简介:Thispaperpresentsanewjointoptimizationmethodforthedesignofsharplinear-phasefinite-impulseresponse(FIR)digitalfilterswhicharesynthesizedbyusingbasicandmultistagefrequency-response-masking(FRM)techniques.Themethodisbasedonabatchback-propagationneuralnetworkalgorithmwithavariablelearningratemode.Weproposethefollowingtwo-stepoptimizationtechniqueinordertoreducethecomplexity.Atthefirststep,aninitialFRMfilterisdesignedbyalternatelyoptimizingthesubfilters.Atthesecondstep,thissolutionisthenusedasastart-upsolutiontofurtheroptimization.Thefurtheroptimizationproblemishighlynonlinearwithrespecttothecoefficientsofallthesubfilters.Therefore,itisdecomposedintoseverallinearneuralnetworkoptimizationproblems.Someexamplesfromtheliteraturearegiven,andtheresultsshowthattheproposedalgorithmcandesignbetterFRMfiltersthanseveralexistingmethods.