臺灣博碩士論文加值系統

此篇論文分別利用兩種方式強健語音辨識系統，一是結合不同模型之優點對倒頻譜參數降維，二是藉由正規化語音特徵統計特性來降低雜訊造成的影響。
第一部分，結合了線性鑑別分析(linear discriminant analysis, LDA)、主成分分析(principal comonents analysis, PCA)、最小錯誤鑑別式(minimum classification error, MCE)三種方法，利用此三種方式之優點，訓練模型來強健語音資料的特徵參數。
第二部分，是以倒頻譜平均消去法(cepstral mean subtraction, CMS)、倒頻譜平均值與辨異數正規化法(cepstral mean and variance normalization, CMVN)為基礎，使用只含語音部分的特徵來建構碼簿(codebook)，同時對每個碼字(codeword)給予權重值(weight)。另外也整合上述之碼簿式與整段式得到的統計資料，發展出組合式特徵參數正規法。最後利用高斯混合模型(gaussian mixture model, GMM) 建立語音模型。最後再加以比較、討論。

This thesis is investigated in two ways to reach enhance the recognition rates respectively, one is to combine different modes which have different advantages, and the other is cepstral statistics normalization techniques to reduce noise effect.
The first part of thesis combines Linear Discriminant Analysis,(LDA), Principal Comonents Analysis,(PCA), and Minimum Classification Error,(MCE) to train model and reach robust the speech feature.
The second part of thesis uses cepstral mean subtraction (CMS), cepstral mean and variance normalization (CMVN) as basic. And normalization techniques use the characteristic parameter which is speech-only to construct codebook. Normalization techniques give weight for every codebook. Then codebook and utterance information are combined in cepstral statistics for normalization. And template matching employs Gaussian Mixture Model, (GMM). Finally, this way compares and discusses the results which are tested in several variable background noises form different conditions.

Acknowledgments i
Abstract in Chinese ii
Abstract in English iii
Contents…. iv
List of figures vi
List of tables vii
Chapter 1 Introduction 1
1.1 Motivation 1
1.2 Introduction of feature extraction 1
1.3 Introduction of hybrid model 2
1.4 Introduction of normalization techniques 2
1.5 Thesis organization 3
Chapter 2 Modified two-dimensional Cepstrum 4
2.1 Introduction 4
2.2 Pre-emphasis 7
2.3 Hamming Window 7
2.4 Discrete Fourier Transform (DFT) 8
2.5 Mel Filter Banks 8
2.6 IIR High Pass Filter and Full Wave Rectifier 9
2.7 Logarithm Transform 10
2.8 Discrete Cosine Transform (DCT) 10
2.9 IDFT and Take Real Part Coefficients 11
Chapter 3 Recognition algorithm 12
3.1 Gaussian Mixture Model [14] 12
3.2 Expectation Maximization (EM) [15] 13
3.3 Speech Recognition 14
Chapter 4 Triple hybrid model for PCA, LDA and MCE 16
4.1 Principal Component Analysis (PCA) 16
4.2 Linear Discriminant Analysis (LDA) 18
4.3 Minimum Classification Error (MCE) 20
4.4 Combination of PCA, LDA and MCE 21
Chapter 5 Cepstral Statistics Normalization Techniques 23
5.1 Utterance-based CMS (U-CMS) and Utterance-based CMVN (U-CMVN) 23
5.2 Codebook-based CMS (C-CMS) and Codebook-based CMVN (C-CMVN) 24
5.3 Hybrid-based cepstral statistics normalization techniques [16] 27
Chapter 6 Experiments and Results 29
6.1 System Specification 29
6.2 The Experiment Results of Hybrid Model (LDA+PCA+MCE) 31
6.3 The Experiment Results of Normalization Techniques 34
6.3.1 The Experiment Results of baseline 34
6.3.2 The Experiment Results of Utterance-based Normalization Techniques 35
6.3.3 The Experiment Results of Codebook-based Normalization Techniques 37
6.3.4 The Experiment Results of Hybrid-based Normalization Techniques 38
6.3.5 The Experiment Results of U-CMS, C-CMS and CU-CMS 40
6.3.6 The Experiment Results of U-CMVN, C-CMVN and CU-CMVN 40
Chapter 7 Conclusion and Future Work 42
Bibliography 43

[1]Wu, G.D. “Chip Design of LPC-ceqstrum for Speech Recognition,” 6th IEEE/ACIS International Conference on Computer and Information Science (ICIS 2007)
[2]Ricotti, L.P. “Multitapering and a wavelet variant of MFCC in speech recognition,” IEEE Proc.-Vis. Image Signal Process, Vol. 152, No. 1 pp. 29-35, February 2005
[3]Pai, H. F. “Two-Dimensional Cepstral Distance Measure for Speech Recognition,” Acoustics, Speech, and Signal Processing, 1993. ICASSP-93., 1993 IEEE International Conference on Vol. 2, pp. 672 - 675 1993
[4]Lin, C.T. “Ga-Bassed Noisy Speech Recognition Using Two-Dimensional Cepstrum,” IEEE Transactions on Speech and Audio Processing, Vol. 8, No. 6, pp. 664-675, Nov. 2000
[5]R.C.T Lee, "Application of principal Compenent Analysis to Multikey Ssearching," IEEE Transactions on Software Engineering, Vol. SE-2, No. 3,pp.185-193, 1976.
[6]J-W. Hung, "Optimization of temporal filters for constructing robust features in speech recognition," IEEE Transactions on Audio, Speech, and Language Processing, Vol.14, No. 3, pp.808-832,May 2006.
[7]J. Ye "A Two-Stage Linear Discriminant Analysis via QR-Decomposition," IEEE Transactions on Analysis and Machine Intelligence, Vol. 27, pp.929-941,No.6, June 2005.
[8]Fu, M.Q. Juang, B.H. Zhou, J.L. and Soong, F.K. “Generalization of the minimum classification error (MCE) training based on maximizing generalized posterior probability (GPP),” in Proceedings of the International Conference on Speech and Language Processing, (Pittsburgh, PA), Sep. 2006.
[9]S. Furui, “Cepstral Analysis Technique For Automatic Speaker Verification ,”IEEE Trans. Acoust. , Speech, Signal Process. ,vol. ASSP-29, no. 2, Apr. 1981.
[10]C.-P.Chen, K.Filality, and J.A. Bilmes, “Frontend Post-processing And Backend Model Enhancement On The Aurora 2.0/3.0 databases,” in Proc. Int. Conf. Spoken Lang. Process.(ICSLP), 2002.
[11]F. Hilger and H. Ney, “Quantile Based Histogram Equalization For Noise Robust Speech Recognition,”in Eur Conf. Speech Communication and Technology (Eurospeech), 2001.
[12]C.-W. Hsu and L.-S.Lee, “Higher Order Cepstral Moment Normalization(HOCMN) For Robust Speech Recognition,”in Proc. IIEEE Int. Conf. Acoust., Speech, Signal Process.(ICASSP), 2004.
[13]Pai, H. F. “Two-Dimensional Cepstrum and its application to Mandarin speech recognition,” Ph.D. dissertation, Inst. Elect. Eng.,Nat. TsingHua Univ., Taiwan, R.O.C., 1993.
[14]Reynolds, D.A. “Robust text-independent speaker identification using Gaussian mixture speaker models,” IEEE Transactions on Speech and Audio Processing, Vol. 3, No. 1, pp. 72-83, Jan. 1995.
[15]Pai, H. F. “Two-Dimensional Cepstrum and its application to Mandarin speech recognition,” Ph.D. dissertation, Inst. Elect. Eng.,Nat. TsingHua Univ., Taiwan, R.O.C., 1993.
[16]J-W. Hung, “Incorporating Codebook and Utterance Information in Cepstral Statistics Normalization Techniques for Robust Recognition in Additive Noise Environments,” IEEE Signal letters, Vol.16, No.6, June 2009.