基于深度学习与并行计算的语音增强系统-计算机技术专业论文.docxVIP

基于深度学习与并行计算的语音增强系统提出了一种利用多模型方法解决系统泛化能力不足问题的框架。针对该框架提出了三种能同时兼顾增强效果和匹配效果的模型。除此之外，本文在模型训练过程中使用GPU进行加速运算，并提出了“动态噪声叠加”算法克服单GPU显存瓶颈。经过试验，系统平均综合PESQ性能提升能够达到0．3274，最高PESQ性能提升达到0．3647。关键词深度神经网络，并行计算，语音增强，理想二值掩蔽II万方数据内蒙古大学硕士学位论文ASpeechEnhancementSystembasedonDeepLearningandParallelComputingAbstractInrecentyears，withthedevelopmentofcomputingcapabilityandtheadventofoptimizationalgorithms，real-timespeechrecognitionbecomespracticablegradually．Althoughtheaccuracyofclosetalkrecognitionhasbeengreatlyimproved，therearestillsomeobstaclestodailyuse．Itismainlybecausespeechisusuallyinterferedbynoisesinrealenvironment。Whilespeechenhancementtechniquesaimtoprocessnoise-mixedaudiosignalandreconstructcleanspeechdatafromthemixture．SpeechEnhancement，alsoknownasSpeechSeparationandSpeechDe-noise，isatechniquetoextractvaluablespeechsignalfromnoise-mixedvoicedataandalsosuppressnoisesignal．Asaresult，ittendstoincreaseSignal-NoiseRatio(SNR)ofaninputsignalandimprovespeechintelligibilitySpeechEnhancementisanessentialpreprocessofvariousapplications．Itisvaluableonresearchandindustry．E．g．automaticspeechrecognition(ASR)，hearing-aidItelecommunication．Therefore，effectivenessofSpeechEnhancementhasaprofoundimpactontheseworks．Therearemanyalgorithmsforspeechenhancement。Mostofthemfocusondigitalsignalprocessing(DSP)methodwithmanycomplexmodelsandalgorithms．Inthisstudy，1willutilizedeepneuralnetwork(DNN)withparallelacceleratedcomputingtosolvetheproblemofspeechenhancement．Deepneuralnetworkextendstheartificialneuralnetwork(ANN)withmorehiddenlayers)mappingtheinputfeaturetothetarget．Withhugedata，DNNcangetridofmanualfeatureextractionanduserawdatadirectlytoIII万方数据基于深度学习与并行计算的语音增强系统extractthevaluablerepresentationfromtheinputautomatically．Thiskindoffeaturehasastrongadvantageofrepresentationlearning．OnepracticalproblemofDNNisthehighcomputationalcost，whichlimitsitsapplication。Howeveritisgraduallybeingacceptedbythedevebopmentofcomputingcapability．Withtheemergenceofgeneralpurposegraphicprocessunit(GPGPU)，parallelcomputingdeveloped