基于LeNet-RES的室內(nèi)聲源區(qū)域定位算法

打開(kāi)文本圖片集

關(guān)鍵詞：室內(nèi)聲源定位；麥克風(fēng)陣列；神經(jīng)網(wǎng)絡(luò)；相位變換加權(quán)廣義互相關(guān)函數(shù)；信號(hào)處理；房間脈沖響應(yīng)中圖分類(lèi)號(hào)：TN912-34；TP183 文獻(xiàn)標(biāo)識(shí)碼：A 文章編號(hào)：1004-373X（2025）10-0020-05

Abstract：Undertheconditionsoflowsignal-to-noiseratio（SNR）andhighreverberation，indorsoundsourceregional localizationbecomesmorechalenging.Inordertosolvethisproblem，aneuralnetworkLeNet-RESisdesigned，whichuses residualblockstoimproveLeNet，therebyimprovingtheperformanceofthenetwork.Thedatasetforindoorsoundsourceis obtainedbysimulating theroomimpulseresponseofan8-araycuboid microphonearay.Thesignalreceivedbythemicrophone is processed intoframes，ndthegeneralizedcrosscorelationPHAse transformation （GCC-PHAT）betweeneachframesignalsis calculated.Thisfunctionisarrngedintotwo-dimensionaldataasinputfeatures.Thefialnetworkmodelistrainedbytakigthe rompartitionarealabelasthenetworkoutput.Intheexperiment，thepositioningaccracyofthetwoneuralnetworkswas tested whenthenumberofroompartitionswas8and16，respectively.Theresultsshow thatunder thesameSNRconditions，the accuracyofLeNet-RES-16is 81.33% when thenumber of room partitionsis16 and the reverberation time is O.6s，which is 23% higherthanthatofLeNet-16;underthesamereverberationconditions，theaccuracyofLeNet-RES-16is84.16%whenthe numberofpartionsis16andtheSNRisO，whichis29%higherthanthatofLeNet-16.Theregionallocalizationperformanceof LeNet-RES is better than that of LeNet under various SNR and various reverberation times.

Keywords：indoorsound sourcelocalization；microphonearray；neural network；generalizedcrosscorelation PHAse transformation; signal processing;room impulse response

0 引言

傳統(tǒng)室內(nèi)聲源定位采用麥克風(fēng)陣列信號(hào)處理算法來(lái)實(shí)現(xiàn)，被廣泛應(yīng)用于智能家居、安防監(jiān)控、大型會(huì)議、車(chē)載系統(tǒng)等多個(gè)領(lǐng)域[1-5]。（剩余7727字）