Webpage:https://sites.google.com/site/deeplearningsourceseparation/
ExperimentsMIR-1Kexperiment(singingvoiceseparation)Trainingcode:codes/mir1k/train_mir1k_demo.m
Demo
Downloadatrainedmodelhttps://www.ifp.illinois.edu/~huang146/DNN_separation/model_400.matPutthemodelatcodes/mir1k/demoandgotothefolderRun:codes/mir1k/demo/run_test_single_model.mTIMITexperiment(speechseparation)Trainingcode:codes/timit/train_timit_demo.mandcodes/timit/train_timit_demo_mini_clip.m
Demo
Downloadatrainedmodelhttps://www.ifp.illinois.edu/~huang146/DNN_separation/timit_model_70.matPutthemodelatcodes/timit/demoandgotothefolderRun:codes/timit/demo/run_test_single_model.mTSPexperiment(speechseparation)Trainingcode:codes/TSP/train_TSP_demo_mini_clip.m
Demo
Downloadatrainedmodelhttps://www.ifp.illinois.edu/~huang146/DNN_separation/TSP_model_RNN1_win1_h300_l2_r0_64ms_1000000_softabs_linearout_RELU_logmel_trn0_c1e-10_c0.001_bsz100000_miter10_bf50_c0_d0_7650.matPutthemodelatcodes/TSP/demoandgotothefolderRunthedemocodeatcodes/TSP/demo/run_test_single_model.mDenosingexperimentPutoriginalFCJF0,FDAW0',FDML0,FECD0,'FETB0','FJSP0','FKFB0','FMEM0','FSAH0','FSJK1','FSMA0','FTBR0','FVFB0''FVMH0oftheoriginalTIMITdataundercodes/denoising/Data/timit/
Trainingcode:codes/denoising/train_denoising_demo.m
Demo
Downloadatrainedmodelhttps://www.ifp.illinois.edu/~huang146/DNN_separation/denoising_model_870.matPutthemodelatcodes/denoising/demoandgotothefolderRunthedemocodeatcodes/denoising/demo/run_test_single_model.mDependenciesThepackageismodifiedbasedonrnn-speech-denoising
ThesoftwaredependsonMarkSchmidt'sminFuncpackageforconvexoptimization.
Additionally,wehaveincludedMarkHasegawa-Johnson'sHTKwriteandreadfunctionsthatareusedtohandletheMFCCfiles.
WeuseHTKforcomputingfeatures(MFCC,logmel)(HCopy).
Weusesignalprocessingfunctionsfromlabrosa.
WeuseBSSEvaltoolboxVersion2.0,3.0forevaluation.
WeuseMIR-1Kforsingingvoiceseparationtask.
WeuseTSPforspeechseparationtask.
Workonyourdata:Totrythecodesonyourdata,seemir1k,TSPsettings-putyourdataintocodes/mir1k/Wavfileorcodes/TSP/Data/accordingly.
Lookattheunittestparametersbelowcodes/mir1k/train_mir1k_demo.m,codes/TSP/train_TSP_demo_mini_clip.m(withminibatchlbfgs,gradientclipping)
Tunetheparametersonthedevsetandchecktheresults.
ReferenceP.-S.Huang,M.Kim,M.Hasegawa-Johnson,P.Smaragdis,"JointOptimizationofMasksandDeepRecurrentNeuralNetworksforMonauralSourceSeparation",IEEE/ACMTransactionsonAudio,Speech,andLanguageProcessing,vol.23,no.12,pp.2136–2147,Dec.2015
P.-S.Huang,M.Kim,M.Hasegawa-Johnson,P.Smaragdis,"Singing-VoiceSeparationFromMonauralRecordingsUsingDeepRecurrentNeuralNetworks,"inInternationalSocietyforMusicInformationRetrievalConference(ISMIR)2014.
P.-S.Huang,M.Kim,M.Hasegawa-Johnson,P.Smaragdis,"DeepLearningforMonauralSpeechSeparation,"inIEEEInternationalConferenceonAcoustic,SpeechandSignalProcessing2014.
NotesThecodesaretestedusingMATLABR2015a
RelatedImplementationssource_separaton_ml_jeju
评论