ѧ
JOURNAL OF SOFTWARE
2000Vol.11No.2P.271-276



ģ͵Զѡ㷨
⻪첨̩
ժҪʶ,ξõѡѵ,ʹ串ǾܶһǳҪ.ͳѵϲֹѡٽмͲķ,˷Ա֤ѡĸ.һԶشӴģϿѡϵ㷨,㷨ʹѡϸǼ,ܱ֤ѵӺ㹻,ʹѵݲϡ,Ϊѵȷɿģʹ˼ʵĻ.
ؼʣʶ,ģѵ,,.
ţTP391
Automatic Corpus Selecting Algorithm Based on Triphone Models
WU HuaXU BoHUANG Tai-yi
(National Laboratory of Pattern Recognition Institute of Automation The Chinese Academy of Sciences Beijing 100080)
AbstractIn speech recognition,the selection of training corpus for robust acoustic modeling which can cover almost all phone phenomena is very important.Traditionally,corpus is selected manually first,and then tested and supplemented,which can't provide sufficient coverage of samples for various statistical modeling methods.An algorithm for automatically selecting the training samples from large-scale text corpus is developed in this paper.This algorithm can not only cover almost all phone phenomena but also ensure to include ideal samples of triphones or class-triphones and ensure enough data for training,which makes it possible to train acoustic model reliably.
Key wordsSpeech recognition,model training,triphone,class-triphone.
ʶ,ģͻģ͵ѵ,һѡϵ.ЩѡӦеԺʹ,Ϲģһ,ӦʹϾܸе,ʹѵݲϡ.ѵ,ǵʵ,ѡѵ뾭Ƶѵ,ܴ5%ʶʵĲ,ɼ൱Ҫ.
Ҫ,һЩʶоλԴ˸˸߶ȵ,йѧԺԶо1995Ƶcosdicϣ1,֤㷨˴һѡ;廪ѧҲ1995깹һ699ʵĴʱ2.Щѡ񷽷µȱ: ѡصڴһ,ʶĶΪ,ƱӰʶЧ; ʱҪںڼĶԪ,,ÿͬʱǰӵӰ,ǿҵЭͬ.,1996ѧԺоڡ863ݿ⡱Ŀ֧,ձѡ1560,ڴϵĻɲͬظ¼һݿ.ڸѡĲֹ˹ɵ,ʱ,޷ɸģѵ,ˣֵһϵѵģ͵³.
ĸĸɵ,ĸɸ,ĸɵԪ򸴺Ԫ.һ,֮ڲǿҵЭͬ,ĸĸķڵصӰ.ĸĸģͽЭͬ,ӶӰʶȷ.,ʶҪģ,ĸĸ֮ڵصӰ.ѡʱ,ӦʹѡϸǺе,ݷصӷ,Խϡ⣨ѵģʱ,뱣֤ÿгֵĴ10Σܻ֤ģ͵׼ȷ,ִʱ,ͳΪϡ裩.
һԻľģΪӦñġĳӶӺӵĸǹ׵ΪݵԶѡϵķ,˷Ա֤ǽ϶ľѡ,ܽϡ.
ȽӺӵĸ,Ȼѡϵı׼㷨,һЩʵ.
1 ʶӺ
ĵĸĸ,ĸ21,ĸ9ĸ13ĸ15ĸɣ3.ǵڽṹ,ʶпģ(磬ģ)ʱ,һеĸĸΪĽģλ.ڿıʱ,ֻĸβĸͷӰ.ģͿдX-Y-Zʽ,Xڵĸβ,Yĸĸ,Zڵĸͷ.ǵѵеľģ,ǵõ8͵ģ:(1) βĸͷ;(2) ĸĸĸ;(3) ĸĸͷ;(4) βĸĸ;(5) βĸͷ;(6) ĸͷ;(7) ĸĸ;(8) βĸ+.ӵɼ1.
Table 1Constituents of triphones
1 ӵĿ

XYZ
InitialsFinals
a,o,e,er,i,
i1,i2,u,v,n,ng,
silence,21 (initials)b,d,g,p,t,k,
z,zh,j,c,ch,q,f,
s,sh,x,h,m,n,l,ra,o,e,i,i1,i2,u,v,er,
ai,ei,ao,ou,ia,ie,ua,uo,ve,
iao,iou,uai,uei,an,ian,uan,van,en,in,
un,vn,ang,iang,uang,eng,ing,ong,ionga,o,e,er,i.
i1,i2,u,v,silence,
21 (initials)

˵:(1) Xеĸβnĸеġnͬ,ǰ߲ܴYλ,ڴ˷ֿ,
(2) i1zi,ci.siеi,i2zhi,chi,shiеi,
(3) erķȽ,ǲereΪͷrΪβʽ,
(4) silence,
(5) ТΪĸ,Ϊĸ.
ͨ,еĸĸ,21ĸһĸ37ĸ,Ӧ814,ʵֻ410.ͬ,ӵ,ĸص㼰ĸĹص,ڲΪ22295,ʱΪ23745.
ѡʱ,ǵӵϱȽ϶,ϡ.,ģ͵ѵ,һ϶»¶ϵľ෽ϡ,,ûھķ4.,ѡʱ,ֿδӵĸʵ,˶ζѧϵݵĸʵķԵΪҪ.ڴ,ǰXZеĸ䷢λͷصΪ83.
(1) :b,d,g;
(2) :p,t,k;
(3) :z,zh,j;
(4) :c,ch,q;
(5) :f,s,sh;
(6) :m,n;
(7) :l;
(8) ͨ:r.
ڲΪ10374,ʱΪ11378.ɼ,Ӹ˴Լһ.
2 ѵԶ㷨
ǵĿǴӴģѡһľΪѵ,Ϻصµ˼.
(1) ǵںڼǿЭͬ,ΪʵطӳЭͬ,ǲᵽģ.
(2) ѵģʱڵϡ,ϽᵽXZеĸ䷢λͷصз,ѧѵеľ˼ǳǺ.
(3) ǿǵʶ,һλ,,ѡĶʵеľ.
(4) ȫԶ˹Ԥѡ,,ʹ߿԰ԼԸѡ,ͽѡٵ,ϵ³.
(5) ԭ,ľȱѡ.
(6) ȫ渲ԭ,֤ѡѡδֹӵľ.
(7) ĳӰгƵʵ͵,˾ӿɱظѡ.
ѡϷȱ˼,һԪģΪġʵѡӵȫԶѡ㷨.ΪϵԶ㷨,Ǳһַ,ܹһӳĶ.,һ,˺ܹ֤ᵽԭȫ渲ԭ.
2.1 
ʵÿӵĵ÷,Ŀ: ԭȫ渲ԭ; ϡ,֤ѡнķһķΧ.ѡ,,һΪӱ,еӼѡгֵĴ.һΪӱ,еӼѡгֵĴ.ʵʽ:
ĳӦмΪ,
ӱеļҲΪ,
score+=W3ʽеġ+C++еĸֵŵͬ
,
score+=W2
ĳڱеļ,
ӦСĳһަ1,
score+=W1+W4Ӧ;
Ӧަ1Сĳһަ2,
score+=W1+W5Ӧ;
,
score+=W1;
ӵĵ÷:
score+=scorenum;
scoreһ,ÿӶӺӵĸǹ.Խ,ѡеĿԽ.Wi,i=1,2,3,4,5ǸʵȷȨֵ,W3W2W1,W4W5,21,numΪĿ.
ϲǸʵȷ,W3W2W1,֤ܹǾܶ;21ʹҵӲϡ,ʹӵķһΧ.
2.2 ѡ㷨
㷨ӺΪġΪֶε㷨,ʵʽ:
(1) ʵϽԤ,̵̫̫ľ˵,ĸдŵľӹ˵.
(2) Դע.
(3) ʼ.ӺΪ,ôѡļĴnΪ.
(4) 빩ѡ.
(5) ӺÿӼƷ.
(6) ÿӰյ÷ָߵͽ.
(7) ѡ÷ߵľ,Ĵnļ1.
(8) ѡӺ,Ӧı.
(9) nǷﵽԤֵ,,ת(10),,ת(5),ѭ.
(10) .
3 㷨׼ʵ
һ㷨Ƿɿ,Ҫһֶκͱ׼.ѡϵ㷨ǷɿԼscoreW3,W2,W1ͦ1,2Ƿǡ,ѱѡеǵĿĿԼӦƽϡ״Ϊָ.Ȼ,ҵӺĿԽֵ࣬ƽԽ,㷨Խ.
3.1 ʵõԭʼ
ԭʼǴձѡ,5ļ,ÿļԤĴСԼΪ5Mֽ.ѡԭʼ϶ĸ,2,е1еġ˵ұߵͳƽǰļͳƽĻϵó.5ļӺֱΪ20 3779 505,ĸʷֱΪ85.7%83.5%.ͳ,Ƿδӵ,Ǻгָʼ΢С.Ҫ,ֱǡĸĸͷ͡βĸͷ,ֱռ1/32/3.ǰںбΪĸںĸڵ,ҴͷͬʱֿɵΪһ,硰o;Ҫ3ĸڵ,硰a+a+e֮ӣʵ.ķжϣԭʼǾдԺձԵ.
Table 2Initial information of corpus
2 ϵԭʼϢ

total number of included triphonestotal number of included class-triphones
File 1179298447
File 2193049034
File 3198629265
File 4201679398
File 5203779505

ٰ,ڰ.
3.2 ʵ
ʵ.ʵ1ǶԲϽڽܵ㷨ϡ³;ʵ2ǵ㷨ѡ1560ԼȾѡϵķıȽ.ʵ1Ľ3,ʵ2Ľ4.
Table 3Information of selected corpus
3 ѡϵϢ

File 1File 2File 3File 4File 5
total number of selected triphones1792919304198622016720377
total number of selected class-triphones84479034926593989505
average frequency of each triphone21.940.463.582.798.5
average frequency of each class-triphone43.762.781.5101.3119.0
the number of triphones appearing more than ten times64018898100851110212098
the number of class-triphones appearing more than ten times51255958703174607790

ٱѡ,ڱѡ,
ӳֵƽ,ӳֵƽ,
ѡгִ10ε,ѡгִ10ε.
Table 4Comparison of training corpus
4 ѵϱȽ

100000
sentences selected
automatically1560*64 sentences
selected by the
Institute of Linguistics100000 sentences
selected through
equidistant sampling
total number of selected triphones20377730717049
total number of selected class-triphones950543748078
average frequency of each triphone98.55.1*64120.0
average frequency of each class-triphone119.08.6*64269.8
variance of selected triphones450.111.5*64471.0
variance of selected class-triphones673.616.4*64764.6

Զѡ100 000,о1560*64,۵Ⱦѡ100 000,
ܱѡݱѡӳֵƽ
ӳֵƽ౻ѡӵķᱻѡӵķ.
5ļ,ÿļ20000,һ100000,ԼΪ3Mֽ.ÿļֻ20 000,ǰļõӺϢΪѡ.ʵʱ,W1,W2,W3ֱΪ20.0,18.0,2.0;1,2ֱΪ20.0,1.0.ӱ2ͱ3пԿ,㷨԰ԭʼаӺӶѡ,еӺϢ.,űѡϵ,ִ10εӺĿ֮,ռıռıʹԼ30%,˵ܴ󽵵ϡĳ̶,ҲǰӷҪԭ.ΪеĽϡķɲͬ¼ͬ,ǰᵽ³,ǵ㷨ܹȫԶѡɲͬͬ,ϵһƫ.
4ǵ㷨ѡϺ1560ӼȾıȽ.Ϊ˱֤ȽϵĿ,ѵ64ͬ¼ƶ,γ99840,ѡ100000൱.,ǰߵϲɶ¼,ǵӺӵǲ,ֱֻ30%38%,Ⱦֱ𸲸Ӻ71.8%71.0%,ǵϿɸ85.7%83.5%.ͬʱ,ǰߵķӶ,ɢҲ֮.ͬ,ȾķҲȽϴ,óķԱȽС,ÿгֵĴԱȽƽ.ɼǵ㷨ڸʺͷ涼вɱԽ.
4  
һԶѡȡѵϵ֮Чķ,ܸѡа,ܽϡ.ʶĽǶȿ,ѵԽࡢԽȫ,ʶЧԽ,ǵ㷨ѡ֤ʶϵ,ǵҲԽ,ϡҲܵõõĽ.
ĿоõȻѧ(No.69835030).
߼飺⻪,Ů,1974,ʿ,ҪоΪʶ,ȻԴ.
첨,1966,ʿ,оԱ,ʿʦ,ҪоΪʶ,˻ͨ.
̩,1934,оԱ,ʿʦ,ҪоΪϳɡʶ𼰴,
Ϣ.
ߵλ⻪(йѧԺԶоģʽʶصʵҡ 100080)
첨(йѧԺԶоģʽʶصʵҡ 100080)
̩(йѧԺԶоģʽʶصʵҡ 100080)
οף
1Qu Fei,Huang Tai-yi,Zhang Xi-jun.Acoustic library design for mandarin Chinese.In:Proceedings of the 4th National Conference on Man-Machine Speech Communication.Beijing,1996.337341(,̩,ϣ.ۺ.:4ȫ˻ͨѶѧļ.,1996.337341)
2Sun Jia-song,Wang Zuo-ying,Wang Xia et al.Construction of the lexicon for continuous acoustic model training.In:Wu Quan-yuan,Gao Wen eds.Proceeding of the Improvement of Intelligence Computer Interface and Application'95.Beijing:Beijing University Publisher,1995.116121(,Ӣ,.ѵʱĹ.:ȪԴ,ı.'95ܼӿӦýչ.:廪ѧ,1995.116121)
3Lin Tao,Wang Li-jia.A Course Book of Phonetics.Beijing:Beijing University Press,1991(,.ѧ̳.:ѧ,1991)
4Gao Sheng,Xu Bo,Huang Tai-yi.Class-triphone acoustic modeling based on decision tree for mandarin continuous speech recognition.In:Li Hai-zhou,Guan Cun-tai eds.Proceedings of the '98 International Symposium on Chinese Spoken Language Processing Symposium Processing.Singapore:The Chinese and Oriental Language Information Processing Society,1998.4448
ոڣ1999-01-07
޸ڣ1999-03-18
