Զѧ
ACTA AUTOMATICA SINICA
2000Vol.26No.3P.369-372



K-MEANS㷨ﾳ
ʸ
󪳶ᡡֱ١Ԭ
ժҪоʶﾳʸ.k-means(k-ֵ)㷨һصҶĸﾳʵֶģ͵ĻܶȵŻ.ʵk-means㷨ﾳʸõƽֲܶȱȼ򵥺ϲҶõƽֲܶ4%10%.
ؼʡʶﾳʸk-means㷨ܶȵŻ.
CONTEXT-DEPENDENT VECTOR QUANTIZATION
BASED ON K-MEANS ALGORITHM
XU XiaobinDING FengLIN BiqinYUAN Baozong
(Institute of Information Science,Northern Jiaotong University,Beijing100044)
AbstractAn approach based on k-means algorithm to implement context-dependent vector quantization for continuous speech recognition is presented.With the approach,individual phonetic context of each leaf in a decision tree can be moved from one leaf to another,in order to optimize the mixture density given by that tree.Experimental results given in this paper demonstrate that the average likelihood given by this approach based on k-means algorithm is four to ten percent higher than that obtained with the previous method of optimizing mixture densities by merging leaves of decision trees.
Key wordsContinuous speech recognition, context-dependent vector quantization, k-means algorithm, optimization of mixture density
1
ĿǰʶϵͳУΪʹģܹ׼ȷڸﾳеķ뷢ﾳصʸ12ȷģеĻܶ.ÿܶȵķ뷢ﾳ޹صģͶѵViterbi alignmentãﾳ޹ģʹӦڸַﾳеѵѵѵɿԽϸߣܹ׼ȷؽÿһ֡ѵݷֱ𻮷ֵӦ״̬תƣ.
ǣлٶȺʹ洢ռСģͿɵĻܶȷĿ.ˣͬﾳķݱҪͬһϷʾ.Ҫʸ֮󣬾ֻٵļƬĿǰһ㲻10ֹڵ㣨ҶӣΪܶȵķ.
׼ȷزЩҶӻϷﾳʸҪо.
ĸһָĽﾳʸصڣ1)Ȳֵﾳ⣬ֵﾳʸԶ2)ÿһҶӵÿһﾳѵݼϿԵϲҶУԻöģ͵ĻܶȵŻ.
2здڵ
ﾳʸ.һַ2ֹ趨ڵĶֵ⣨ﾳ⣩ʸУЩֵⲻɸı.ʸȱ趨ĶֵԽϴ󣬻Ӱࣨܶȣľȷ̶.һַ1෴Ҫֹ趨ڵĶֵ⣬ʸУڵĶֵԸı䣬ӶʵֶԻܶȵŻ.
ȻһַȱǣĶֵĸıͨϲƬݷֲܶƵҶӣϷʵֵ.ÿƬҶӿܰﾳӶʹÿκϲǶﾳݵġںϡֻĳһﾳѵݺϲҶ.ʱܻΪ޷ÿһﾳһĵܻŵĻܶ.
3k-meansﾳʸ
ڶķһֻk-means㷨ﾳʸ.ַУģ1иﾳʸ㷨Ϊصÿһ״̬תƣһþkƬҶӵľȻk-means㷨ЩҶķﾳ䷢ݣϽеֱõһŵĻܶ.
3.1ѵﾳ޹ģ
ķһﾳʸ֮ǰӦصȫﾳѵѵһɢܶȵHMM.HMMʸ뱾С210256֮.ʸһ׵ϵ֣ЩײӵԼʸ뱾.⣬ÿһHMMҲӵԼʸ뱾ɵ׼ײֵĶ뱾ɣ.ÿһģ͵˽ṹڸصﾳģͣҲﾳ޹ģͣͼ1ʾ3״̬ÿһ״̬һǰתƺһת.״ֱ̬ԼĻܶ.


ͼ1ģ͵˽ṹ
ﾳ޹ģ͵ĳʼѵֹзֵ.ڳʼѵɺøģͣﾳ޹ģͣȥԶзֺͱעڸﾳеѵݣԶзֺͱעһֱ̣Ϊֹ.
ɢܶȵģﾳ޹أHMMѵ֮øصɢHMMڸַﾳеѵViterbi alignmentЩѵݻֵģ͵ĸ״̬.ȻΪÿһģ͵ÿһ״̬һþ.ĸڵӵͬһ״̬ȫﾳеѵ.
3.2k-means㷨ĻܶȵŻ
ڵķѺͶԻܶȷŻԶֵеǸݽڵķ棨goodness-of-splitƽֲܶɵ.
ҪѵģP0ĳһڵnһﾳqQ(QnйP0صļϣﾳĴ𰸵Ĳֵǡ򡰷񡱣nѳӽڵ㣨ӽڵnlӽڵnr.Ӧķmn(q).nnlnrķֲܶȷֱPnPlPr.GaussianֲʾΪ
Px(y)=N(y,x,x).(1)
ʽx=n,l,r;x,xֱnnlnrľֵʸͣԽǣЭ.
mn(q)Ըʽ1
(2)
ʽyl,kyr,k,yn,kֱǻֵnl,nr,nѵ;Nl,NrNnӦʸĿqʾʹnѵﾳ.
ͨ棬ԴģP0ĸڵﾳУõǰKﾳѵݼϣΪĳʼҶӣܶȵĸ.ȷʼҶӵĿģΪܶȷŻ̲õk-means㷨k-means㷨ܵӣں̶ܴȡڳʼķ༯ϵģ3.
ȷĳʼҶ֮k-meansԻܶȽŻԸҶеķﾳеʹƽֲܶȴﵽ.ʽƽֲܶ
(3)
ʽyi,jǵiƬҶӣܶȵĵiϷĵjj=1,2,,Niѵ
4ʵ
ѵ6˵˵1000.ÿһ֡ѵݣʸ12melϵ40ms80msϵһײ֣20msϵĶײ֣Լ˲ʱһײ֣logƵʼһײ֣֡25ms֡Ƽ10msƵ11kHz.
ǰʹģ1ݵ㷨ɳʼﾳʸǰKﾳΪĳʼҶӣһҶǶﾳĻ.ȷʼҶ֮k-means㷨Ҷеﾳһؽеʹƽֲܶȴﵽ.ʵ1ʾ.
1ֱk-means㷨Ҷӡںϡ㷨õƽֲܶ

ҶӣϷĿƽlogֲܶ
k-means㷨Ĺֵƽlogֲܶ
ҶںϣĹֵ
2-2.24-2.31
3-1.51-1.57
4-0.89-0.99

 иҶֱΪ234ʱƽֲܶȣֵ.е2иǲk-means㷨ԾиҶӵķﾳһеʱõƽֲܶȣõŻܶȷ3иģ1ݵ㷨ƵľҶӽСںϡʱõƽֲܶ.Կk-means㷨ԷﾳһؽеõƽֲܶȾڽƵľҶӽСںϡõƽֲܶ.ɴ˿ɼk-means㷨ԷﾳһؽеԸЧضԻܶȷŻӶõ׼ȷĻܶ.
лоõģʽʶʵһ̩оԱָ͹ڴ˱ʾл.
ȻѧĿ.
󡡱ͨѧźϢרҵʿ.ҪоΪʶ˻Իϵͳ.
ᡡͨѧźϢרҵʿ.ҪоΪźŴȻԴ.
ֱ١19431965ҵڱͨѧΪͨѧϢѧо.ڴźŴĽѧͿй.
Ԭڡڣʿʦ.йѧźŴֻϯ.ڴͼ󡢼ӾȻԴģ̬Ϣʵо.
󣨱ͨѧϢѧо100044E-mail:xiaobinxu@263.net
ᣨͨѧϢѧо100044
ֱ٣ͨѧϢѧо100044
ԬڣͨѧϢѧо100044
ο
1ݡBahl L R,Souza P V De et al. Context dependent vector quantization for continuous speech recognition. In: Proc.1993 IEEE Int. Conf. on Acoustic, Speech, and Signal Processing, Minneapolis MN May. 1993
2ݡHwang M,Huang X,Alleva F A. Predicting unseen triphones with senones.IEEE Trans. Speech and Audio Processing. 1996,4(6)412419
3ݡTou J T.Pattern Recognition PrinciplesMassachusetts: Addison-Wesley Publishing Company, 1974
ոڡ1997-12-23
޸ڡ1999-11-20
