생몰정보
소속
직위
직업
활동분야
주기
서지
국회도서관 서비스 이용에 대한 안내를 해드립니다.
검색결과 (전체 1건)
원문 있는 자료 (1) 열기
원문 아이콘이 없는 경우 국회도서관 방문 시 책자로 이용 가능
목차보기더보기
title page
Abstract
Contents
List of Abbreviations 13
I. Introduction 14
1.1. Motivations of the Research 14
1.2. Contribution 16
1.3. Outline 17
II. Compensation for Acoustic Mismatch in Automatic Speech Recognition 19
2.1. Acoustic Mismatch in ASR 19
2.1.1. Noisy Speech Model in Signal Space 19
2.1.2. Noisy Speech Model in Feature Space 21
2.2. Speech Enhancement Techniques in Signal Space 24
2.2.1. Wiener Filter 25
2.2.2. MMSE-STSA 28
2.2.3. MMSE-LSA 34
2.3. Feature Compensation Techniques in Feature Space 37
2.3.1. CMN 37
2.3.2. CMVN 39
2.3.3. HEQ 39
2.3.4. Comparison of CMN, CMVN, and HEQ 43
III. Class-Based Histogram Equalization 45
3.1. Introduction 45
3.2. Hard Class-Based Histogram Equalization 49
3.3. Soft Class-Based Histogram Equalization 52
3.4. Class-Tying Technique 56
3.4.1. Hard-Class-Based Histogram Equalization with Class-Tying 57
3.4.2. Soft -Class-Based Histogram Equalization with Class-Tying 60
IV. Experiments and Results 63
4.1. Experimental Framework 63
4.1.1. Speech Database 64
4.1.2. Baseline Feature Extraction 66
4.1.3. Baseline Speech Recognition System 71
4.1.4. Performance Evaluation Measures 71
4.2. Experimental Results for Speech Recognition 73
4.2.1. Transformation by Acoustic Mismatch 73
4.2.2. Baseline MFCC features 79
4.2.3./4.2.2. MMSE-LSA-Based Speech Enhancement Technique 81
4.2.4./4.2.3. Conventional Feature Compensation Techniques 85
4.2.5./4.2.4. Class-Based Histogram Equalization Techniques 91
4.2.6./4.2.5. Utilizing the Class-Tying technique 97
4.2.7//4.2.6. Incorporation of the MMSE-LSA Technique 101
4.2.8./4.2.7. Comparative Performance Analysis 111
V. Conclusion 122
국문요약 126
References 128
Acknowledgement 133
Curriculum Vitae 135
Table 2.1. Comparisons of the feature space-based compensation methods. 44
Table 4.1. The structure of the Aurora-2 database. 65
Table 4.2. The average word error rates using the baseline MFCC feature. 80
Table 4.3. The average word error rates using the MMSE-LSA method. 84
Table 4.4. The average word error rates using the CMN method. 88
Table 4.5. The average word error rates using the CMVN method. 89
Table 4.6. The average word error rates using the HEQ method. 90
Table 4.7. The average word error rates using the hard-CHEQ method. 95
Table 4.8. The average word error rates using the soft-CHEQ method. 96
Table 4.9. The average word error rates using the hard-CHEQ method with the classtying technique. 99
Table 4.10. The average word error rates using the soft-CHEQ method with the classtying technique. 100
Table 4.11. The average word error rates using the MMSE-LSA and CMN methods. 104
Table 4.12. The average word error rates using the MMSE-LSA and CMVN methods. 105
Table 4.13. The average word error rates using the MMSE-LSA and HEQ methods. 106
Table 4.14. The average word error rates using the MMSE-LSA and hard-CHEQ methods. 107
Table 4.15. The average word error rates using the MMSE-LSA and soft-CHEQ methods. 108
Table 4.16. The average word error rates using the MMSE-LSA and hard-CHEQ methods with the class-tying technique. 109
Table 4.17. The average word error rates using the MMSE-LSA and soft-CHEQ methods with the class-tying technique. 110
Figure 2.1. Acoustic mismatch between training and test environments of ASR. 20
Figure 2.2. Noisy speech model in signal space. 21
Figure 2.3. The basic idea of HEQ-based feature compensation. 41
Figure 3.1. The side effect of conventional HEQ in case of nonmonotonic transformation by the acoustic mismatch. 45
Figure 3.2. The side effect of conventional HEQ in case of the distribution mismatch between the training and test data. 47
Figure 3.3. The concept of CHEQ-based feature compensation. 48
Figure 3.4. Block diagram of the hard-CHEQ method for feature compensation. 51
Figure 3.5. Block diagram of the soft-CHEQ method for feature compensation. 55
Figure 3.6. Block diagram of the tied-hard-CHEQ method for feature compensation. 59
Figure 3.7. Block diagram of the tied-soft-CHEQ method for feature compensation. 62
Figure 4.1. Block diagram of the MFCC-based feature extraction by the Aurora Group 70
Figure 4.2. Transformation caused by the acoustic mismatch between clean and noisy… 76
Figure 4.3. Transformation caused by the acoustic mismatch between clean and noisy … 77
Figure 4.4. Transformation caused by the acoustic mismatch between clean and noisy … 78
Figure 4.5. The procedure of robust speech recognition employing the MMSE-LSA based speech enhancement technique. 83
Figure 4.6. Recognition results of hard or soft-CHEQ with or without class-tying for… 94
Figure 4.7. Incorporation of the MMSE-LSA technique to the feature compensation methods in reducing the acoustic mismatch for robust speech recognition. 103
Figure 4.8. The average word error rates by various compensation methods. 115
Figure 4.9. The error reductions by various compensation methods over MFCC and conventional HEQ. 115
Figure 4.10. The average word error rates by various compensation methods on test set A. 116
Figure 4.11. The error reductions by various compensation methods over MFCC and conventional HEQ on test set A. 116
Figure 4.12. The average word error rates by various compensation methods on test set B. 117
Figure 4.13. The error reductions by various compensation methods over MFCC and conventional HEQ on test set B. 117
Figure 4.14. The average word error rates by various compensation methods on test set C. 118
Figure 4.15. The error reductions by various compensation methods over MFCC and conventional HEQ on test set C. 118
Figure 4.16. The error reductions by various compensation methods over MFCC on each noise of test set A. 119
Figure 4.17. The error reductions by various compensation methods over HEQ on each noise of test set A. 119
Figure 4.18. The error reductions by various compensation methods over MFCC on each noise of test set B. 120
Figure 4.19. The error reductions by various compensation methods over HEQ on each noise of test set B. 120
Figure 4.20. The error reductions by various compensation methods over MFCC on each noise of test set C. 121
Figure 4.21. The error reductions by various compensation methods over HEQ on each noise of test set C. 121
원문구축 및 2018년 이후 자료는 524호에서 직접 열람하십시요.
도서위치안내: / 서가번호:
우편복사 목록담기를 완료하였습니다.
* 표시는 필수사항 입니다.
* 주의: 국회도서관 이용자 모두에게 공유서재로 서비스 됩니다.
저장 되었습니다.
로그인을 하시려면 아이디와 비밀번호를 입력해주세요. 모바일 간편 열람증으로 입실한 경우 회원가입을 해야합니다.
공용 PC이므로 한번 더 로그인 해 주시기 바랍니다.
아이디 또는 비밀번호를 확인해주세요