欢迎来到天天文库
浏览记录
ID:36468236
大小:2.44 MB
页数:85页
时间:2019-05-11
《非特定人连续语音识别的理论、分析和实验》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、浙江工业大学硕士学位论文非特定人连续语音识别的理论、分析和实验姓名:徐春燕申请学位级别:硕士专业:检测技术与自动化装置指导教师:王守觉;曹文明20041201别的新算法。本文比较了在不同数量的建模样本下,HNI模型和高维空间点覆盖神经网络模型对数字音节的建模能力,实验结果表明,样本数量较少时,高维空间点覆盖神经网络模型的识别率明显高于HMM模型,高维空间点覆盖神经网络模型能较合理地描述样本在特征空间的分布,对样本的信息利用率高。高维空间点覆盖神经网络通过对原构网样本和识别错误的样本的重新学习,能明显提高识别率,表明高维空间点覆盖神经网络有较强
2、的学习能力;通过对识别错误的样本单独构建的高维空间点覆盖神经网络,与原来的高维空间点覆盖神经网络之间进行有机结合,识别率显著提高,证明高维空间点覆盖神经网络有较强的即时学习能力以适应环境。这种不同模型之间易于集成的能力可以用在多语种多方言的语音识别系统综合集成方面,为语音识别系统的广泛应用提供了光辉的发展前景。关键词连续语音识别单词音节隐马尔可夫模型高维空间几何高维空间点覆盖神经网络ⅡTHETHEORY,ANALYSISANDEXPERIMENTSABoUTSPEAKER-INDEPENDENTCoNTINoUSSPEECHRECoGNlTI
3、oNABSTRACTSpeaker-independentcontinuousmandarinfigurespeechrecognitionsystemhaswideapplicationinmanyfieldssuchastelephonespeechgateway,remotecontrolofhomeappliances,industrycontrolandinformationinquirysystem.Ithastheoreticalandpracticalimportanceinbothfewvocabularyandlargev
4、ocabularyspeechrecognitionsystems.Owingtoitscharacteristicssuchasshorterpronunciation,easierconfusionandmanymoredialects,mandarinfigurespeechrecognitionismoredifficultthanEnglishfigurespeechrecognitionandbecomesachallengingtopicinthespeechrecognitionfield.HiddenMarkovModel(
5、HMM)whichisbasedonstatisticsbecomesthemainspeechrecognitiontechniquenOWdaysbecauseitsstrongcapabilityinmodelingthedynamictimesequenceandflexibilityintheselectionoftheparameters,structuresandtrainingtechniques·However,InHMMnotonlyneedsagreatdealofsamplestotrainitsmodelwhichm
6、akesthesamplecollectionandtrainingabigworkandhaslowusagerateofsampleinformation,butalsocan’ttrainthesampleoflowprobabilitysuffic.ientlywhichmightleadtofalserecognition·Thispaperismainlytostudythefiguresyllablemodelingforspeaker-independentcontinuousmandarinfigurespeech·Anew
7、human..computercombinedfiguresyllablefeatureextractiontechniqueisproposed.Thetrainingsetandthetestsetoffiguresyllableareextractedfromthecontinuousfigurespeech.Guidedbythehi曲dimensionalspacecoveringtheorythehighdimensionalspacecoveringneuralnetworkforeveryclassoffiguresyllab
8、leisconstructed.Therecognitionrateoftheltestsetreachesmorethan97%.Thismodelisappli
此文档下载收益归作者所有