о뷢չ
JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT
1999Vol.36No.6P.664-667



ںִɹѧϰеӦ
͢ۿġ
ժҪɹҪƵ仯ɣĿǰɹͨ˹õһЩ.Ϊѧϰִʵɹнɹǹڴз֮ӳϵӶʹѧϰӳ.нַṹʵ.
ؼʣϳɣɹʱ
ţTP18
NEURAL NETWORK IN LEARNING MANDARIN 
TWO-WORD PHRASES FULL PITCH VARIATION PATTERNS
ZHU Ting-ShaoGAO Wen
(Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100080)
Charles LING
(Department of Computer Science,University of Western Ontario London, Ontario, Canada, N6A 5B7)
AbstractProsodic model includes pitch model and duration model,and these models being used are described qualitatively.In order to learn the prosodic patterns from Chinese two-word phrases,the patterns are regarded as the mapping between the isolate syllable and the same one when it is spoken in phrases,and the mapping can be learned by using neural network.In the paper here,the architecture of learning neural networks is introduced,and the results are also given.
Key wordsspeech synthesis, prosodic rules, time wrapping, neural network
1
Ϊ˺ϳɸȻȸȵƵɹ򣬶ɹܽӦòöķ.ɹҪڵʱֲߵı仯ı仯ʵͣٵȼ棬ߺı仯ȻȵӰ1.ڼã2Ժģʽܶо,һЩԵ.ïӵȣ3ͨͨѧõͨǰڵߵĹϵ.1ô洢14ֵڵģʽ22ִģʽкϳʱ޸ģĸʱ޸ģĸʱղͬʱϵе.Ŀǰߺ仯ɴǸѧоóһЩԵǶɹĲֶ.
ɹҪڸµߺı仯ڹڴзµĻƵкܴ仯ɹԿǶԹںʹзһӳϵ.˼·ʹ洫ѧϰں͸ڴзʱĻƵʱӳϵѵõֱӼĻƵ.
ϵͳɴġͬʱҲǼ򵥵ĴԪԪ㷺ػӶγɵĸϵͳ.Ϊֹ˺ܶģ.洫 (back propagation
,BP)һֶṹӳ磬һЩصʹΪĿǰӦ㷺һģͣ4.
2ṹ
ִʳֵĸǶԺִʵƵ仯ɷֱѧϰ.һִлƵܴ仯ǰڣڵı仯ǰߵ.
ʹѧϰһѧϰִǰڻƵı仯һѧϰڻƵı仯ɣʱѧϰУѵΪƽֱѧϰ.6磬ԷΪ飬3ѧϰǰڵĻƵ仯3ںйعɵѧϰ.ÿ3УһѧϰִжӦڵĻƵ仯ɣֱѧϰӦΪƽµʱ仯.ͼ1˽ṹͼ.


ͼ1Ƶɵѵʹ
ͬһڵĻڹºڴʱͬģĳЩϳ㷨ԣԻΪλкϳɣ˻ı仯Ҳͷӳʱı仯.ʱ仯ɵѧϰͨѧϰƵı仯ʵ.
ʹɹѧϰҪǽڹԼڴзǰ仯һӳϵӶʹBPѧϰӳ.ͼ1(a)ǽѵ̣ͼ1(b)˽вԺʹõ.
3Ԥ
ݿõϳϿCoSS-1ǡ֧Ŀ廪ѧϵйѧԺѧоѧԺоͬɵ.CoSS-11268еں640ִʣͬ¼ѹκŲ迹.Ų迹ǽ˵˵񶯲μ¼αȽԸ.ĿǰȻкܶƵ㷽ȷȶߣ5.ͨŲĻעԵõ׼ȷĻֵ.ݿȡĻƵȡֵΧΪ50260˵ѵݵȡֵΧӦ-1.0+1.0˶ԱעõĻƵݱб׼.
ΪѵҪ󣬱뽫ȲͬĸѵеĻƵͬһȣڳΪPitchLenĻƵPitchΪWrapLen.ĻƵ0j<WrapLenPitchLenеӦλӦΪLoc=j * PitchLen/WrapLen˹ĻƵjλõĻƵӦΪPitchӦĻƵLocӦֵ.Ϊ˼ƵLocλõֵǲ˲ֵķҳLoc4Ƶ㣬Ȼʹ4ڵֵͨõLocӦֵ.
4ɹѧϰ
4.1Ƶ仯ɵѧϰ
ڻƵ仯ѧϰΪ3㣺㡢㣬֮ȫ.ѧϰȡ0.25ʹSigmoidΪڵΪ100ʹֱѧϰִѵǰںͺڵĻƵ仯.
ڵаڵʱĻƵֵǰںﹲ4ʹ8ڵʾǰڵǰ4ڵʾ1ڵ4ڵʾ2ڵ磺01001000ʾ1Ϊ2ڵΪһ.ڵΪִзʱĻƵֵЩƵֵǾȶִʽз֡Ƶעʱͱ׼õ.ڵڵĸĳȶͬȡйڵĻƽֵΪĻƵݵĹȣִǰڵĻڵĻڵƽֵΪӦĽڵ.
4.2ʱ仯ɵѧϰ
ѧϰʱ仯ɵƵ仯ѧϰƣΪ3㣺㡢㣬֮ʵȫ.ѧϰȡ0.25SigmoidΪڵΪ20.
һ4ѧϰʱı仯ɣѧϰѵǰڵʱı仯ƽڵʱ仯ʹһѧϰʹһ磻ҲʹͬѵͲ.ִеǰڣڹʱĻƵҰִʵǰڵĸ.ڵΪִзʱĻƵ.
5ʵ
Ϊ˶ѵвԣǴCoSS-1ѡһΪʵݣֱڻƵʱѵͲԣвûаѵ.
1еͳǻڹִڵĻƵֵ뾭ѵõĻƵֵĻϽеģСֵͳ߲ľֵеģͳԭʼƵƵĻֱӽеģߵıȽϿֱӿѧϰЧ.2ʱĲʵ.
һЩԽõĲͨʾеĻƵ߿ڻƵעܴ.ѵݽƽѧϰЧнһ.ʱԽԿƽѧϰǺ룬ĲԽʵݱȽϽӽ.
Ϊ˸ز鿴ѧϰЧǻһʾ򣬽ԽͼʾʵʻƵֵõĻƵֱֵ۵ʾԺؿߵǺϳ̶.ƪޣڴ˾Ͳʾʵ.
6
ɹںϳɸǳҪĿǰʹõһЩɹһЩѧĶ.ǲ粢ھִеĻƵʱ仯ɽѧϰȡ˽Ϻõѧϰ.ûѧϰķйȡʹöԴģݽдΪܣͨʹ֪ʶֵ㷨ԸõѧϰЧ.ϣûѧϰõȫϵͳ仯ɣߺϳ.
1Ƶʵ

ƴMaxMinMean
ǰchuang1
shan440-1.4772064.376950
jin4
zhan4130-0.24121042.273769
liu2
xie4180-4.52561615.284344
gong1
hui41502.65648310.662880
wang2
pai2902.0161557.791716
zhi4
xun260-0.0010932.958961
bing3
yao4140-3.63853923.004646
bao3
xian370-0.52798410.082108
la1
suo3700.1686617.902920
xiao4
you31904.45394624.817568

2ʱʵ

ƴڴֵԽ
ǰƽla1suo35150
cao1yan33549
chong1ren42941
chuan1ma33650
a1pian43642
tu2an43130
chuo4hao43229
shu4e22729
cuo4shou33735
fu3ai43530
ƽta1yang15051
xiang1Yue17551
sui2jun14151
qie4sheng14043
ban4yin15052
can2ren34440
du1fu33336
wei2ao44237
yi2zhi43540
fu2fu23939


ĿõйѧԺ˼ƻĿ.
߼飺͢ۿУ197112ʿооΪھݿϳ.
ģУ19564ڣʿʦйѧԺ㼼оҡƻܼר鳤(ϯר)оԺԺҪоΪ˹ܺͶý弼ڼӾģʽʶͼýѹ.
壬УôԴѧϵڣоΪѧϰھ.
ߵλ͢ۿйѧԺ㼼о100080
ģйѧԺ㼼о100080
壨Դѧѧϵô󰲴ʡN6A 5B7 
οף
1.ȸȻȺתϵͳо۲ʿģݣйѧԺѧо,1995
(Chu Min. Research on Chinese TTS system with high intelligibility and naturalness (in Chinese). Institute of Acoustics, Chinese Academy of Sciences, Beijing,1995)
2ڼ.ͨе仯.й,1982,(6):439449
(Wu Zongji. The tone variation in Mandarin. Chinese Grammar (in Chinese), 1982, (6): 439449)
3ï,վ,.ĳʵ.,1984,(1)
(Lin Maocan, Yan Jingzhu, Sun Guohua. Experiment of the normal accent in Beijing dialect.Dialect (in Chinese). 1984,(1))
4ΰ.˹ԭӦ.պѧ,1995
(Wang Wei. Principle of Artificial Neural NetworkPrimer and Implementation (in Chinese). Beijing:Beijing University of Aeronautics and Astronautics Press, 1995)
5о,ٻ.źִ.ӹҵ,1990
(Yang Xingjun Chi Huisheng et al. Speech Signal Digital Processing (in Chinese). Beijing:Publishing House of Electronic Industry,1990)
ոڣ1998-07-14
޸ڣ1998-11-02
