计算机工程与应用
計算機工程與應用
계산궤공정여응용
COMPUTER ENGINEERING AND APPLICATIONS
2014年
20期
215-217,243
,共4页
吕小听%李昕%屈燕琴%胡晨
呂小聽%李昕%屈燕琴%鬍晨
려소은%리흔%굴연금%호신
稀疏表征%高斯混合模型(GMM)均值超向量%超完备字典%最大后验(MAP)算法
稀疏錶徵%高斯混閤模型(GMM)均值超嚮量%超完備字典%最大後驗(MAP)算法
희소표정%고사혼합모형(GMM)균치초향량%초완비자전%최대후험(MAP)산법
sparse representation%Gaussian Mixture Model(GMM)supervectors%over-complete dictionary%Maximum-A-Posteriori(MAP)algorithm
近年来,随着信号的稀疏性理论越来越受到人们的关注,稀疏表征分类器也作为一种新型的分类算法被应用到话者识别系统中。该模型的基本思想是:只要超完备字典足够大,任意待测样本都能够用超完备字典进行线性表示。基于信号的稀疏性理论,未知话者的向量系数,即稀疏解可以通过L1范数最小化获取。超完备字典则可视为语音特征向量在高斯混合模型-通用背景模型(GMM-UBM)上进行MAP自适应而得到的大型数据库。采用稀疏表征模型作为话者辨认的分类方法,基于TIMIT语料库的实验结果表明,所采用的话者辨认方法,能够大大提高说话人识别系统的性能。
近年來,隨著信號的稀疏性理論越來越受到人們的關註,稀疏錶徵分類器也作為一種新型的分類算法被應用到話者識彆繫統中。該模型的基本思想是:隻要超完備字典足夠大,任意待測樣本都能夠用超完備字典進行線性錶示。基于信號的稀疏性理論,未知話者的嚮量繫數,即稀疏解可以通過L1範數最小化穫取。超完備字典則可視為語音特徵嚮量在高斯混閤模型-通用揹景模型(GMM-UBM)上進行MAP自適應而得到的大型數據庫。採用稀疏錶徵模型作為話者辨認的分類方法,基于TIMIT語料庫的實驗結果錶明,所採用的話者辨認方法,能夠大大提高說話人識彆繫統的性能。
근년래,수착신호적희소성이론월래월수도인문적관주,희소표정분류기야작위일충신형적분류산법피응용도화자식별계통중。해모형적기본사상시:지요초완비자전족구대,임의대측양본도능구용초완비자전진행선성표시。기우신호적희소성이론,미지화자적향량계수,즉희소해가이통과L1범수최소화획취。초완비자전칙가시위어음특정향량재고사혼합모형-통용배경모형(GMM-UBM)상진행MAP자괄응이득도적대형수거고。채용희소표정모형작위화자변인적분류방법,기우TIMIT어료고적실험결과표명,소채용적화자변인방법,능구대대제고설화인식별계통적성능。
The signal sparse theory has received more and more attentions in recent years. Sparse representation, a new classification method for speaker identification has been applied into the speaker identification system. The main idea based on this new approach is that an unknown test utterance can be represented as a linear combination of the training database while the training patterns are sufficient. According to the sparse theory, the coefficients of unknown test utterances corresponding to the class index of test models could be obtained by L1-norm minimization. Over-complete dictionary could be developed by adapting speech features to Gaussian Mixture Model-Universal Background Model(GMM-UBM) using Maximum-A-Posteriori(MAP) adaptation. This paper makes use of the sparse representation model for speaker identification, and the experiments conducted on TIMIT acoustic-phonetic continuous speech corpus show that the perfor-mance of the proposed method consistently outperforms the state of art speaker identification classifiers.