哈尔滨工业大学硕士毕业论文模板

Size: px

Start display at page:

Download "哈尔滨工业大学硕士毕业论文模板"

诚呆叶
7 years ago
Views:

1 I -

2 Abstract Under the development of information globalization, people need text-tospeech technique more and more. This has made outstanding improvement in research and development of text-to-speech technique. Although the quality of synthesized speech has been accepted, the naturalness is still dissatisfying. This paper starts with part of the front end technique, which are text-to-syllable, detection of prosody phrase boundaries and speech stress determining, in order to solve the pronunciation errors and the improper handling in rhythm, pause, stress during text to speech. So the synthesized speech of text-to-speech system can be made more natural. After analyzing the shortcomings of dealing with polyphones in methods based on rule and on statistics, we introduce a comparatively self-contained flow, which are three steps: look in the polyphone list, fit the part of speech rules and compute semantic similarities. This paper describes How-Net in detail and regards it as the setting of semantic computation and deduction in order to compute the word similarity and the context window similarity. The experiment results show that this kind of method for dealing with polyphones has many virtues: using semantic computing can find the semantic similar word of the unknown pronunciation word, and then get the pronunciation of polyphones in it from its similar word. So this method can improve not only the accuracy of determining polyphone pronunciations greatly but also the accuracy of text-tosyllable in text-to-speech system. Though some researchers have got some good experiment results in detection of prosody phrase boundaries, their methods are all very complex. In this paper we build a self-learning system for detection rules of prosody phrase boundaries using the transformation-based error-driven learning, which has succeeded in parsing. After building a large corpus that has 5725 text and speech sentences, the self-learning system retrieves a large amount of rules. Not only retrieving of these rules is easy to use but also the accuracy of automatic detection of prosody phrase boundaries can be raised when using these rules. In - II -

3 Abstract order to reduce the burden brought by the large quantity of rules in TTS system, this paper uses tree to organize rules. The automatic detection of prosody phrase boundaries has been used well in real system. Whether the word stress formulation exists in Chinese mandarin has not got the last word, but research of the Chinese mandarin stress is still very important to text-to-speech system. So this paper starts with the principle of producing and apperceive of prosody word stress and analyzes the rules of two syllable words stress. Then using the same self-learning system as detection of prosody phrase boundaries to get rules of prosody phrases stress. But the experiment results are not good enough to be used in the real system. We must research it further in the future. stress Keywords polyphone; semantic similarity; prosody phrase; prosody word; - III-

5 ... I Abstract... II

6 TTS

7 [ 1] [ 2], [ 3] [ 4 ] - 1 -

8 [ 5] [ 6] 1.2 [ 7] 10 30ms LPC - 2 -

9 1 1Text-To-Speech2 Concept-To-Speech3Intention-To-Speech TTS 1-1 [ 8] 1-1 Figure 1-1 Text-to-Speech System Sketch Map

10 Holmes1973Klatt/ 1980 DEC DECtalk1987 Klatt/ PSOLA1990 PSOLA LPC PSOLA [ 9] LPCPSOLA 863 KX-PSOLA1993, 1995 TH_SPEECH(1993) KDTALK 1995PSOLA [ 10] - 4 -

11 [ 11] TTS 1.4 TD-PSOLA 1-2 TTS Figure 1-2 the Frame of TTS Demo System

12 PSOLA 1-2 1) 2) 3) 1.5 TTS TTS - 6 -

13 [ 12] TTS 10 6 [ 13] TTS 85% Table 2-1 Examples of Polyphone Word hao3 shi4 hao4 shi4 cheng2 qing1 deng4 qing1 zhuan3 xiang4 zhuan4 xiang4 TTS TTS - 7 -

14 TTS TTS WordNetMindNetFrameNet [ 14] [ 15] 2.2 HowNet [ 16]

15 a) b) c) d) e) - % CPU f) - & g) -? h) //- * - 9 -

16 i) //- $ j) - * k) l) m) - n) - o) - p) # 2-1 Figure 2-1 Example of How-Net Relation - 10-

18 W_X= E_X= G_X= DEF= W_XG_XE_X X X C E DEF - 12-

19 2 NO.= W_C= G_C=V E_C=~~~~~ W_E=play G_E=V E_E= DEF=exercise,sport E_C ~ W_C DEF ENTITY entity thing space mental physical event organization inanimate animate shape plant AnimalHuman bacteria human animal 2-2ENTITY Figure 2-2 the Frame of ENTITY Sememe Category Tree - 13-

20 EVENT 2-2 ENTITY entity thing space human entity entity Y_C 5 [ 17] W_C= G_C=V E_C=~~~ Y_C=pa2 DEF= pa2pa

21 [ 18] a b DISTANCE-ATOM(a,b)=a b a b b a Figure 2-3 Sketch Map of the Distance of Sememes DISTANCE-ATOM(a,b)= DISTANCE-ATOM(a,b)=3 a b SIM-ATOM(a,b) (1DISTANCE-ATOM(a,b)/ (2)) 100 (2-1) I Relations(I) K Item-Relation- Atoms(I,K) IJ SIM-WORD-ITEM(I, J)= Max a Item-Relation-Atom(I,K) K Relation(I) b Item-Relation-Atom(J,K) Relation( I) SIM ATOM ( a, b) (2-2) Relation( I) I

22 A Item-Word(A) A,B SIM-WINDOW-ITEM(A,B)= Max Y Item Word ( B) X Item Word ( A) Item Word(A) SIM WORD ITEM ( X, Y ) Item Word( A) A (2-3) TTS / 2-4 Figure 2-4 Flow of Text to Syllable of Experiment System - 16-

23 2 Y_C I=(W,P,T,X)W P T X 1 0 W_C=W G_CT Y_C X=0 X=0 Y_C X=1 W_C=P G_CT E_C W Y_C W W_C=P G_CT W_C Y_C W W 3 W_C=W G_CT Y_C /d/p/r /ng /ng/vg/d /vz /vg /pcheng2 qing1 deng4 qing1 [ 19] A-1vgV,,V,

24 ,,V,0 W_C= G_C=V W_C= G_C=V E_C=~~~ W_C= G_C=V E_C=~~ X=1 ~ ~ ~~~ Y_C=deng4qing1 DEF= Y_C=cheng2qing1 DEF=,, W_C= G_C=V E_C=~~ Y_C=deng4qing1 DEF= deng4 2-5 Fig ure 2-5 Syllable Determining Flow of in /r /t /vq /ng /vg /ngngn,,n, , ,000 40, A % - 18-

25 2,,N,0 W_C= W_C= G_C=N W_C= G_C=N G_C=N E_C= Y_C= Y_C W_C= G_C=N E_C= Y_C=hang2 W_C= W_C= W_C= W_C= G_C=N G_C=N G_C=N G_C=N E_C= E_C=~ E_C=~ E_C=~ Y_C=hang2 Y_C=hang2 Y_C=xing2 Y_C=hang2 DEF=,, DEF= DEF=,, DEF=, W_C= G_C=N E_C= hang2 Y_C=hang2 2-6 Figure 2-6 Syllable Determining Flow of in

27 [ 20] [ 21] 3.2 [ 22] [ 23] 1. trigram probabilities 2. trigram probabilities + distance 3. trigram probabilities + distance + lookahead L trigrams - 21-

28 L Table 3-1 Open Test Results of Five Methods 1 70% % 3 70% 4 72% 5 69% 90% TTS 1995 Eric Brill [ 24] [ 25] [ 26] - 22-

29 Figure 3-1 Transformation-Based Error-Driven Learning 3-1 [ 27][ 28] 1) 2) 3)

30 % [ 29][ 30] 3.6 [ 31] [ 32]

31 3 [ 33 ] if 0:POS=X->PAUSE=* 0POS=X XPAUSE=* *2 *1 *0 if 0:POS=X&0:LENGTH=Y->PAUSE=* LENGTH=Y Y (1) (2) (3) (4) if 0:POS=X->PAUSE=* if 0:POS=X&0:LENGTH=Y->PAUSE=* if -1:POS=X&0:POS=Y->PAUSE=* if 0:POS=X&1:POS=Y->PAUSE=* if 0:POS=X&-1:POS=Y&-1:LENGTH=Z->PAUSE=* if 0:POS=X&1:POS=Y&1:LENGTH=Z->PAUSE=* if 0:POS=X&0:LENGTH=Y&-1:POS=Z->PAUSE=* if 0:POS=X&0:LENGTH=Y&1:POS=Z->PAUSE=* (5)if 0:LENGTH=X&0:POS=Y&-1:POS=Z&-1:LENGTH=U->PAUSE=* (6) (7) if 0:LENGTH=X&0:POS=Y&1:POS=Z&1:LENGTH=U->PAUSE=* if -1:POS=X&1:POS=Y&0:POS=Z->PAUSE=* if -1:POS=X&1:POS=Y&0:LENGTH=Z&0:POS=U->PAUSE=* if -1:POS=X&1:POS=Y&-1:LENGTH=Z&0:POS=U->PAUSE=* if -1:POS=X&1:POS=Y&1:LENGTH=Z&0:POS=U->PAUSE=* - 25-

32 (8) if-1:pos=x&-1:length=y&1:pos=z&0:length=u&0:pos=v ->PAUSE=* if -1:POS=X&1:POS=Y&0:LENGTH=Z&0:POS=U&1:LENGTH=V->PAUSE=* if -1:POS=X&1:POS=Y&-1:LENGTH=Z&0:POS=U&1:LENGTH=V ->PAUSE=* (9) if-1:pos=x&1:pos=y&0:length=z&0:pos=u&1:length=v& -1:LENGTH=W->PAUSE=* Figure 3-2 Automatic Acquisition Flow of Prosody Phrase Boundary Detection Rules

33 Figure 3-3 Waveform and Phrase Tagging of Example Sentence

34 /m /q /nc /vg /m /q /nc /c /m /q /ng A Table 3-2 Experiment Results of Threshold Choosing 1/3 91% / % /12 92% / % /2 91% /3 87% / Table 3-3 Close and Open Test Results of Different Training Data % 73% % 77.1%

35 3 3.4 TTS if 0:POS=ut->PAUSE=0 if 0:POS=ut&0:LENGTH=2->PAUSE=0 if 0:POS=ut&0:LENGTH=2&1:POS=nd->PAUSE=2 if 0:LENGTH=2&0:POS=ut&1:POS=nd&1:LENGTH=4->PAUSE=2 if 0:POS=ut&1:POS=j->PAUSE=1 if 0:POS=ut&1:POS=vg->PAUSE=2 if 0:POS=ut&1:POS=nd->PAUSE=2 TTS 3-4 0:POS=ut PAUSE=0 0:LENGTH=2 1: POS =j 1: POS =vg 1: POS =nd PAUSE =0 PAUSE =1 PAUSE =2 PAUSE =2 1: POS =nd PAUSE =2 1:LENGTH=4 PAUSE =2 3-4 Figure 3-4 Example of Tree-like Organizing of Rules - 29-

36 TTS TTS a 3-5-bTTS 3-5 Figure 3-5 Comparison between Real Speech and Synthesized Speech - 30-

37 3-31-

38 4 4.1 : [ 34] [ 35] accent stress accentnuclear accent stressaccentprominence [ 36] - 32-

39 4 4.2 Agaach M.C. Sluijter [ 37 ] duration spectaltiltvowel qualityintensity Cutler Butterfield [ 38] accent - 33-

40 A ) #2 #1 #2 S0 #2 S0 #1 S0 S1#2 S0 S2 S0 S1 S2 2) 5-34-

41 [ 39] % % 5. 80% - 35-

42 % Table 4-1 Stress Rules Based on Part of Speech Alan W Black [ 40] if 0:POS=X->STRESS=* if -1:POS=X&0:POS=Y-> STRESS =* if 0:POS=X&1:POS=Y-> STRESS =* if -1:POS=X&1:POS=Y&0:POS=Z-> STRESS =* STRESSSTRESS 0STRESS / % - 36-

43 4 19% TTS

44 TTS TTS TTS - 38-

46 - 40-

48 - 42-

49 TTS How-Net 369 Table A-2 Common Polyphone List - 43-

50 - 44-

51 - 45-

52 [1] - 46-

53 - 47-

致谢本人自 2008 年 6 月从上海外国语大学毕业之后, 于 2010 年 3 月再次进入上外, 非常有幸成为汉语国际教育专业的研究生回顾三年以来的学习和生活, 顿时感觉这段时间也

致谢本人自 2008 年 6 月从上海外国语大学毕业之后, 于 2010 年 3 月再次进入上外, 非常有幸成为汉语国际教育专业的研究生回顾三年以来的学习和生活, 顿时感觉这段时间也精英汉语和新实用汉语课本的对比研究 The Comparative Study of Jing Ying Chinese and The New Practical Chinese Textbook 专业 : 届别 : 姓名 : 导师 : 汉语国际教育 2013 届王泉玲杨金华 1 致谢本人自 2008 年 6 月从上海外