ѧ
JOURNAL OF SOFTWARE
1999ꡡ1011ڡVol.10No.111999



ʶлڹ鲢зԶ
ż ֣  ս 
ժҪоʵ˺еԶз㷨ڹ鲢зԶ(merging-based syllable detection automaton,MBSDA)㷨.MBSDA㷨˰ĶʱʺͻڵĶ,߶Ƶ֡(1֡֡)źŽС鲢(merging),γɡ鲢ƶ(merged similar segment,MSS),Ǳ϶ͬһڵͬ״̬.ЩMSSһ״̬ġзԶ(syllable detection automaton,SDA),ڵзֵ.ÿȷзֶڸķΧ(range of syllable number,RSN)ҲMBSDA㷨.
ؼʡз,鲢,зԶĸ,ڸΧ.
ͼšTP391
Merging-based Syllable Detection Automaton in Continuous Chinese Speech Recognition
ZHANG Ji-yong ZHENG Fang DU Shu SONG Zhan-jiang XU Ming-xing
(Speech Laboratory Department of Computer Science and Technology Tsinghua University Beijing 100084)
Abstract In this paper, an automatic syllable detection method namely merging-based syllable detection automaton (MBSDA) is studied and implemented. The MBSDA uses a variety of features including the frame energy, the zero crossing rate and the fundamental frequency to merge similar consecutive frames (one or several frames) into one merged similar segment (MSS). The frames in the same MSS are treated as frames of the same state of a phonetic. These MSSs are passed into a syllable detection automaton (SDA) to give the syllable detection results. In addition, the MBSDA gives the range of syllable number (RNS) of each definite detection segment.
Key wordsSyllable detection, merging, syllable detection automaton, vowel feature segment, range of syllable number.
Ŀǰ,ڷضʶ,㷺ʹõѧģHMMģͼĽģ,дͳ֡ͬ㷨1ݻViterbi㷨2ݸ״̬.һĴʷϢģ,ͿѧȻ,ɾʶ3.
ķ,ѧ㷨ⲻݺ.(1) ·ϱը;(2) ״̬дλ.ں̶ܴӰʶϵͳ.ǵʵ,ʹ,зֵ֪,ڵʶҲܴﵽˮƽ4,,ǰһź׼ȷз,Ȼټʶ.,Ŀǰоˮƽľ,κƾʹﵽһǱȽѵ.Щ,ǿԸ׼ȷı߽,Щ֮ı߽ȴѼ.Ĳȡķǰȫȷĵطп,вĵطڵĸΧ(range of syllable number,RSN),ӶϺõؽ.Ϊ,һ֡ϲзԶ(merging-based syllable detection automaton,MBSDA)㷨.ԺǿصΪ㷨ṩ˺ܺõ.
1 㷨ԭ
1.1 ȡ
MBSDA㷨ʹõҪǶʱ֡ʺͻ.ȡͷ֡Ϊλ.ļ5,ڴ˲.
1.2 ֡Ĺ鲢
ͬһصķ,һʱڱȶ;һعɵһʱ,ᷢ仯.,֡ûзȽϴı仯,Ϊͬһص.ڴ,˶֡С鲢(merging)ĸ.
ڽй鲢֮ǰ,ȼǷת(transition).תТתۺ͢ת,ֱ.
ת.ͻȻ仯.ǰ֡()ǰһ֡()Ħ;ǰ֡()ĦСǰһ֡()(ȡ=2).
ת.仯.ǰ֡ǰT֡()ľֵT֡()ľֵ֮ڢתеı仯ϵ(ȡT=3).
ǰ֡Ģת,֡ϡת۱(transition tag,TT).
һ֡֡ûTTǵ鲢ͬһ鲢ƶMSS(merged similar segment).εһҪ,ӳиȶĲ.
1.3 зԶʵ
ͨβֱӸڵзֱ߽,ǹһзԶ(syllable detection automaton,SDA),Զڲ״̬תȷڵзֵ.SDA״̬Ϊ¼:һĸĸαĸβ.SDAе״ֱ̬º.
(SIL)(NOI).ָ͹ʶǺܵ͵ź,Ϣ.Ƚǿʱ,תΪ.
һĸ(SM1)Ͷĸ(SM2).ĸоͳƱ,l,m,n,rĸص,ǽĸ״̬;ʣµĸĸнԵ,ǽǹһĸ.
α(Psudo-SIL).Щڷʱ,ĸĸмһġ͹ȡ.,ڡkai,ĸkĸaiȽϸ,֮Ĺɶεȴܵ,ھ.鵽״̬,Ȼᷢ.Ϊ˶,α״̬.
ĸ(YM)β(YW).ĸһԵص,׼ڡϸߵеĹ.һת䵽һʱ,ĸ͹ʶбȽԵ½,β״̬.ͨʵ,ǵõ˸״̬֮Ĵ¹1ʾ.Զ״̬תͼ1ʾ.

ͼ1 ½зԶ״̬תͼ
1 Զ״̬Ĺϵ

MSSʾһĸĸαĸβ
֡
1005001001000100100005005001001000
ʣ3104050301058030


1.4 SDAз
ǰѾһαڡ򡰹ڡ,ôݺڵص,2еκһ״̬תʱ,˴һڵһڵת.ҲSDAҪзֱǵĵط. 
2 зֱǵ

״̬תԶǰһ״̬Զĵǰ״̬
1ĸβ
2ĸĸ
3βĸĸ

1.5 зֶڸΧ
зֶڸΧ(range of syllable number,RSN)ȷ漸.
ڵnзֶ,ͳƳĸ(vowel feature segment,VFS)ĸ,ΪCVFS(n).ǰзֶεƽڳ(average syllable length,ASL)ֵ,ǳʼ,ɰʽ(֡Ϊλ). 

(1)
l(1)n(-i)=l(1)(n-i),ʾn-iڹƸΪ1зֶγ(֡Ϊλ),ҲǰλǰiڹƸΪ1зֶγ.ʽõǰλǰMڹƸΪ1зֶγƵǰεASLֵ.,Mѡ,̫С,ƶڳȵľֲ仯̫,ȱ;̫,ڸĹȱٶٱ仯ĸ.ͨ,ѡM=10.
ڸRSN֮ǰ,ASLֵһ޲οֵ
CR(n)=ln/ASLn,
(2)
lnʾƵзֶγ.
ôзֶεCVFSֵCRֵ֮,зֶڸȷΪ

(3)
ñĸз㷨,һ,Cmaxֵ3,󲿷Ϊ1.
2 ʵ鼰
2.1 ʵ
ʵʹ.1ݲ863ݿ.ڰİ칫һ²¼,ûĸ.Ǵгȡ55Ů,ÿ¼߸ȡ20.һŮ100ĲԼ.Ϊ˶ԱзֽӰ,ڻǿķ¼˵2.,˵ߵġԡȸź.ݹ100,.ݵĲƵʾΪ16KHz,Ϊ16bit.¼ߵΪÿ150180.֡Ϊ16ms(256).
2.2 з㷨۷
еȷpd͸ΧȷpnµĹʽ.

(4)
ܵзȷpcpdpnͬ.Ϊ˷ӳз㷨гռڵı,гpoutĸ.ֱ.

(5)
2.3 ʵ
ͳƹʽ,ʵз㷨Ĳ,Եõз㷨ڲͬµROC(receiver operating characteristics),ͼ2ʾ,ڸ,MBSDA㷨pcpoutĺ.ǿԸݲͬгMBSDA,ӶõROCϵĲͬ.һ,һʵʹõз㷨Ҫpcӽ100%.

ͼ2 MBSDAз㷨ROC
ROCǿԿ,׷гʻȷʶᵼ㷨ܵ½.һ,гʴΪ90%ʱ,ϵͳ.3гһʵͳƵĽ.
3 ʵͳƽ

(%)еȷ(pd)RSNȷ
(pn)ȷ
(pc)г
(pout)
ݿ
199.298.398.890.3
Ů99.798.599.192.1
298.996.297.689.8
ƽ99.397.798.590.7

ʵĽԿ,,зֵЧܲ.ŮĶԱпԷ,ŮзЧҪ,ԭҪݿеŮҪ.,ͨʵԿ,зֵĽһӰ,㷨ûкܴȵ½,ⷴӳ˱㷨нϺõ³. 
3  
Ǵ㷨ĿԻ㷨ĸӶ,зֵĺʶһ֮Чķ.ֱܱ㷨езȷ,㷨г,ǽŬķ.Կ,ڴ㷨Ļ,ټźŵһЩ,絹ײ,㷨г.гʴﵽܸߵֵʱ,ʵҲǴﵽзڵĿ.,ڱз㷨ԶѾÿһ֡״̬,˶Ը㷨ʵ޸ļʵڵ/з.
߼飺ż,1977,˶ʿ,ҪоΪʶ,źŴ.
֣,1967,ʿ,,ҪоΪ,ʶ,źŴ.
,1976,˶ʿ,ҪоΪźŴ,ʶ.
ս,1972,ʿ,ҪоΪ,ʶ,źŴ.
,1973,ʿ,ҪоΪʶ,źŴ.
ߵλ廪ѧѧ뼼ϵʵ  100084
E-mail: ZhangJY@sp.cs.tsinghua.edu.cn
ͨѶϵ:ż£ 100084,廪ѧѧ뼼ϵʵ
ο
1 Lee C H, Rabiner L R. A frame synchronous network search algorithm for connected word recognition. IEEE Transactions on ASSP, 1989,37(11):16491658
2 Zheng Fang, Chai Hai-xin, Shi Zhi-jie et al. A real-world speech recognition system based on CDCPMs. In: Proceedings of the International Conference on Computer Processing of Oriental Languages (LCCPOL'97). Hong Kong, 1997. 204207
3 ֣,Ĳ¡,ǵ.һתı༭ʵ.:з,ſȱ.5ȫ˻ͨѶѧ(NCMMSC'98)ļ.:ҵѧ,1998.280285
(Zheng Fang, Mu Xiao-long, Xu Ming-xing et al. The implementation of a speech-to-text transition editor. In: Wang Cheng-fa, Zhang Kai et al eds. Proceedings of the '98 National Conference on Man-Machine Speech Communication. Harbin: Harbin Institute of Technology Press, 1998. 280285)
4 ֣,Ļ,.CDCPMʶеӦ.ѧ,1996,7(863ר):6975
(Zheng Fang, Wu Wen-hu, Fang Di-tang. CDCPM with its application to speech recognition. Journal of Software, 1996,7(special issue of 863):6975)
5 о,ٻ.źִ.:ӹҵ,1995
(Yang Xing-jun, Chi Hui-sheng. Speech Signal Digital Processing. Beijing: Publishing House of Electronics Industry, 1995)
1998-10-14յԭ,1998-12-22յ޸ĸ
