基于人耳聽(tīng)覺(jué)特性的語(yǔ)音特征提取研究.pdf

1、Y893327分類(lèi)號(hào)：——密級(jí)：——單位代碼：l逝22學(xué)號(hào)：2Q坦2豎i西北大學(xué)碩士學(xué)位論文作者：麥墊墊指導(dǎo)教師魚(yú)塞堡專(zhuān)業(yè)技術(shù)職務(wù)割奎燃學(xué)科(專(zhuān)業(yè))鹽簋墊鏊鮭量墨鹼答辯日期20066學(xué)位授予日期二oo六年六月AbstractWhichthisarticleformulatesisdividedinto3parts：hearingmaskingequalloudnessandpropertyofcochlearfrequency—divi

2、sionThemaincontributionsasbelow：1)Anapproachcarvingout婦，‘后frequencybandwith，材P，frequencymarkisbroughtforwardandacquiredabark聃lsheetInresearchonthebasisofhearingmasking，aeonceptofmaskingandspeeehrecognitionmethodbytraditi

3、onallyusingmaskingfeatureTraditionalmethodalsoincludesremovingstablenoisesignalfromthespeechbymeansofspectralsubtractionThisarticleformulatestherelationbetween卅eZFrequencyand6口，惦Frequencybandandacquireda婦，彘。耐sheetInexper

4、imenttothissheet，cepstrumexaltationisalsodoneto^I，’CCFirstandseconddifferencewillbedonetoM陽(yáng)Cand0neenergyframesothatacquire39dimensionfeatureparameterstooptimize髓skingfeatureinpickingup朋|，cc2)Unitedweightingof5：^啦andequal

5、10udnesspropertyisbroughtforwardandanewweightingfilterisdesignedaccordingtothenew冊(cè)criterionInresearchbasedonequalloudnesspropertyofhumanauditorysystem，wediscoveredthatcommonequal10udnessweightingdividedspeechsignalintose

6、ctionsbyspectrumandeachofsectionsisweightedrespectively，IftheS撇ofonespeechsectionislow，nodoubt，bythismethod，noisewillimpactmoreonsignalandmakeagainstenhancingrecognitionaccuracyrateThisarticleformulatesthataftersignalpas

7、sedequal10udnessfilter，weightingshouldbedonetotheweightedFrequencybandsbyusing5：^傻Besides，inthisarticle，anewweightingfilterisworkedoutaccordingtothenew冊(cè)equalloudnesscontour3)Addingequal10udnessweightingintosubbandisbroug