資源描述:
《提高耳語音可懂度的非對稱壓縮語音增強方法.pdf》由會員上傳分享,免費在線閱讀,更多相關內(nèi)容在行業(yè)資料-天天文庫。
1、第39卷第4期聲學學報Vo1.39.No.42014年7月ACTAACUSTICAJu1.2014提高耳語音可懂度的非對稱壓縮語音增強方法術周健,2鄭文明3王青云2趙力2(1安徽大學計算智能與信號處理教育部重點實驗室合肥230031)(2東南大學水聲信號處理教育部重點實驗室南京210096)(3東南大學兒童發(fā)展與學習科學教育部重點實驗室南京210096))2012年l2月28日收到2013年5月6日定稿摘要提出兩種基于非對稱代價函數(shù)的耳語音增強算法,將語音增強過程中的放大失真和壓縮失真區(qū)分對待。ModifiedItakura~Sa
2、ito(MIS)算法對放大失真給予更多的懲罰,而Kullback—Leibler(KL)算法則對壓縮失真給予更多的懲罰。實驗結果表明,在低于一6dB的低信噪比情況中,經(jīng)MIS算法增強后的耳語音的可懂度相比傳統(tǒng)算法有顯著提高;而KL算法則獲得了同最小均方誤差語音增強算法近似的可懂度提高效果,證實了耳語音中的放大失真和壓縮失真對于耳語音可懂度的影響并不相同,低信噪比時較大的壓縮失真有助于提高耳語音可懂度,而高信噪比時的壓縮失真對耳語音可懂度影響較小。PACS數(shù):43.72Anasymmetricattenuatedspeechenha
3、ncementapproachforimprovingintel1ibilityofnoisywhisperZHOUJian,。ZHENGWenming。WANGQingyun。ZHAOLi(1KeyLaboratoryofIntelligentComputingandSignalProcessingofMinistryofEducation,AnhuiUniversityHefei230031)(2KeyLaboratoryofUnderwaterAcousticSignalProcessingofMinistryofEduca
4、tion,SoutheastUniversityNanjing210096)(3KeyLaboratoryofChildDevelopmentandLearningScienceofMinistryofEducation,SoutheastUniversityNanjing210096)ReceivedDec.28,2012RevisedMalv6,2013AbstractTwoasymmetriccostfunctionforwhisperedspeechenhancementmethodsareproposed.Thecost
5、oftheamplificationdistortionandtheattenuationdistortionarediferentinbothmethods.TheModifiedItakura-SaitofMIS1distancefunctiongivesmorepenaltiestospeechamplificationdistortionwhiletheKullback—Leibler(KL)divergencefunctiongivesmorepenaltiestospeechattenuationdistortion.
6、TheexperimentalresultsshowthattheMISmethodgainslargerintelligibilityimprovementofthewhisperedspeechthantheconventionalspeechenhancementalgorithmsinmuchlowerSignaltoNoiseRatio(SNR)lessthan一6dB,andtheKLmethodhassimilarintelligibilityimprovementperformancetotheMinimumMea
7、nSquareError(MMSE)speechenhancementmethod.Theresultsconfirmthattheamplificationdistortionandtheattenuationdistortionhavediferentefectsontheintelligibilityoftheenhancedwhisper.Specifically,largerattenuationdistortioncanimprovespeechintelligibilityinlowerSNRconditionand
8、ithasalittleinfluenceonspeechintelligibilityinhighSNRcondition.冰國家自然科學基金(61301295,61231002,61273266,61003131)、安徽省自然科學基金(1308