本論文採用梅爾倒頻譜係數及其差分係數來代表聲學特性，再以Legendre多項式和離散餘弦函數去逼近音調軌跡，而萃取出音調特性的係數。接著對這兩類係數分別建立高斯混合模型，此外，也對各個音框的特徵係數作表徵化，再去建立PPM模型，以模式化連續的表徵之間的語言特性。除了以線外實驗探討適合對國、閩南、客語三種語言作辨識的模型結構外，我們也建造了一個可供線上測試的自動語言辨識系統。在線外測試中，三種語言的平均辨識率可到達97%，而在線上的初步測試裡，辨識率則可達73%之平均值。

In this thesis, different models and model structures are studied to identify among the three languages: Mandarin, Holo, and Hakka. Acoustic features are represented as Mel-frequency cepstrum coefficients (MFCC). In addition, Legendre polynomials and discrete cosine transforms (DCT) are used to approximate the pitch contour of a voice segment. For the two kinds of features, Gauussian mixture models (GMM) are constructed respectively. Also, each frame’s feature vector is tokenized in order to construct a prediction by partial matching (PPM) model to modelize the characteristics embedded in a sequence of tokens. According to the models studied, a practical language identification system has been built. In offline tests, identification success rate can reach 97% in average. In initial online (i.e. inputting with a telephone) tests, the system can obtain a success rate of 73% in average.

[1] K. Li and T. Edwards, “Statistical models for automatic language identification,” Proc. International Conference on Acoustics, Speech, Signal Processing, 1980.
[2] M. A. Zissman, “Comparison of four approaches to automatic language identification,” IEEE Trans. Speech and Audio Processing, vol. 4, pp. 31-44, Jan. 1996.
[3] J. Navrátil, “Spoken language recognition—a step toward multilinguality in speech processing,” IEEE Trans. Speech and Audio Processing, vol. 9, Sep. 2001.
[4] L. Mary, K. S. Rao, and B. Yegnanarayana, “Neural network classifiers for language identification using phonotactic and prosodic features,” International Conference on Intelligent Sensing and Information Processing, 2005.
[5] W. H. Tsai and W. W. Chang, “Discriminative training of Gaussian mixture bigram models with application to Chinese dialect identification,” Speech Communication, vol. 36, pp. 317-326, Mar. 2002.
[6] P. A. Torres-Carrasquillo, D. A. Reynolds, and J. R. Deller Jr., “Language identification using Gaussian mixture model tokenization,” IEEE International Conference on Acoustics, Speech, Signal Processing, May 2002.
[7] C. Y. Lin and H. C. Wang, “Language identification using pitch contour information in the ergodic Markov model,” IEEE International Conference on Acoustics, Speech, Signal Processing, Jun. 2006.
[8] 林俊青，多國語言辨識系統之特徵設計研究，碩士論文，國立中山大學電機工程研究所，2002。
[9] 張智傑，以高斯混合模型表徵器與語言模型為基礎之語言辨認研究，碩士論文，國立清華大學電機工程研究所，2005。
[10] 郭頂益，多國語言辨識系統之設計研究，碩士論文，國立中山大學電機工程研究所，1999。
[11] 林俊宇，應用隱含式語意索引與語言模型於中英夾雜語音之語言鑑別，碩士論文，國立成功大學資訊工程研究所，2001。
[12] University of Cambridge, “Hidden Markov model Toolkit (HTK),” http://htk.eng.cam.ac.uk/.
[13] D. O’Shaughnessy, Speech Communications, 2nd ed., IEEE Press, 2000.
[14] H. Y. Gu, H. F. Chang, and J. H. Wu, “A pitch-contour normalization method following Zhao’s pitch scale and its application,” Conference on Computational Linguistics and Speech Processing (ROCLING), 2004.
[15] H. Y. Gu, S. Y. Sun, and H. F. Chang, “A scoring system for Mandarin tones uttered in disyllabic words,” Conference on Computational Linguistics and Speech Processing (ROCLING), 2006.
[16] A. V. Oppenheim and R. W. Schafter, Descrete-time Signal Processing, 2nd ed., Prentice-Hall, 1999.
[17] Wikipedia, “Windows function,” http://en.wikipedia.org/wiki/Window_function.
[18] Diracdelta.co.uk, “Hamming window,” http://www.diracdelta.co.uk/science/source/h/a/hamming%20window/source.html.
[19] Wikipedia, “Legendre polynomials,” http://en.wikipedia.org/wiki/Legendre_polynomials.
[20] S. J. Leon, Linear Algebra with Applications, 7th ed., Prentice-Hall, 2006.
[21] S. Chakrabarti, Mining the Web, Morgan Kaufman, 2003.
[22] M. Kaufmann, Data Compression, 2nd ed., Morgan Kaufmann, 2000.
[23] W. J. Teahan, “Probability estimation for PPM,” New Zealand Computer Science Research Student Conference, Apr. 1995.
[24] Erik de Castro Lopo, “Secret rabbit code,” “http://www.mega-nerd.com/SRC/index.html.