大词汇连续汉语语音的MLP声学特征的研究.doc

大词汇连续汉语语音的MLP声学特征的研究.doc.doc吻合器经腹联合断流术56例疗效分析
[摘要],PLP作为输入向量的高斯混合模型(GMM)的隐马尔可夫模型(H丽)的经典模型在大词汇连续语音识别系统(LVCSR)已取得了良好识别效果。但针对短时声学特征区分性差的特点,本文提出采用神经网络多层感知器(MLP)产生的两种类型差异特征HATs与 TANDEM代替短时特征,分别训练G酮参数模型。实验结果表明,差异特征的GMH画的LVCSR系统优于传统的短时特征的系统;为了更进一步提高系统识别率,该文又将两种类型差异特征HATs与TANDEM进行复合,构成MLPs特征流重建 GMH丽,系统的错字率(CER)有2%〜%的明显改善。
关键词:多层感知器;差异特征;隐马尔可夫;高斯混合
模型
分类号:TN912文献标识码:A文章编
号:1009-3044 (XX) 13 -3470-02
MLPFeatur esforLargeVocabula ryContinuousMandar inSpeechRecognitio nSystem LVDan-jul,Ch. Plahl2,
(Scienc eDept. ,SouthwestFo restryUniversity,K unming650224,China ;Furlnformatik6-Co
mputerScienceDepar tmentRWTHAachenUni
versity, Aachen5205 6,Germany)
Abstrac t : TypicallyHiddenM arkovModel(HMM)ini
argevocabularycont
itionsystem(LVCSR)
ousticfeaturesvect
utfeaturestotheGau GMM),hasachievedgo ts. However,forthep
rmfeaturesondiscri workmultilayerperc toproducetwotypeso
turesHATsandTANDEM
ures,andrespective gGMMparametermodel
mentalresultsshowt
Rsystembasedondisc
issuperiortothesys
dontheshort-termfe
provethesystemreco
typesofdiscriminat
bineda
retrainGMHMM,thatl
inuousspeechrecogn ,usingshort-termac ors (/PLP)asinp
ssianmixturemodel( odrecognitionresul
ooroftheseshort-te
mination, eptron(MLP)areused
fdiscriminativefea
insteadofshortfeat
ly,thecorrespondin saretrained. Experi
hattheG順MMtheLVCS
riminativefeatures
temtraditionalbase
atures;Tofurtherim gnitionrate,thetwo
ivefeaturesHATsand