透過您的圖書館登入
IP:3.145.60.29
  • 學位論文

外國人華語音調發音偏誤偵測 – 電腦輔助語言學習初步實驗

Mandarin Tone Pronunciation Error Detection – Computer Assisted Language Learning for Foreign Speaker

指導教授 : 廖元甫

摘要


華語是音調 (Tone) 語言,而德語是一種非音調的語言,華語的音調變化會造成語義上的不同,但是德語音調上的變化只會讓句子的語氣改變並不會讓意思不同,因此,華語音調是所有外國語音學習者的共同難點,為了發展華語電腦輔助語言學習系統 (Computer assisted language learning, CALL) 中的音調錯誤偵測模組,我們先蒐集德華雙向語料,並求取參數組合為基礎,再利用Fujisaki Model移除語句的影響,以及z-score的方法作正規化,去訓練Monotone和Tritone的Hidden Markov Model,之後利用Posterior Probability計算出每個音調值做錯誤偵測。 實驗採用TCC300 及我們自己蒐集的華語L2語料庫進行測試,我們採用幾種方法,方法一我們利用TCC300訓練辨認氣,之後將華語L2語料作為測試語料;方法二我們利用華語L2語料取三分之二作訓練辨認器,剩餘的三分之一當作測試語料;這兩種方法我們都各別加上調適語料作調適模型,並觀察調適前與調適後的結果,最後比較所有方法的辨識與錯誤偵測結果。實驗結果顯示,組合參數比起單一參數效果來的好,以TCC300為訓練器辨識華語L2語料,最佳結果得到48.30%;而用華語L2訓練語料為訓練器辨識華語L2測試語料,在還沒加上調適語料前,最佳的結果得到36.27%,但加上調適語料後的結果得到42.16%;錯誤偵測方面,則華語L2作訓練器模型比用TCC300做訓練器的偵測效果來的好,主要是因為德國人在講華語時,音調跟台灣人的音調差太多,而且講話速度也來的比較慢,未來還可以多考慮到講話速度來進一步研究與改進。

並列摘要


Tone pronunciation error detection is the most important issue for a Mandarin Chinese computer-assisted language learning (CALL) system. To develop a good tone error detector, (1) a bi-lingual speech corpus of L2 learners of German and Mandarin is collected, and several (2) Hidden Markov Model (HMM)-based tone (monotone and tri-tone) recognizer and (3) tone verification modules are built and compared. The experimental results using TCC300 as the training and L2 Mandarin speech corpus as the test set shows that (1) tone recognition error rates, (2) verification errors (EER). The best recognition error rate is 48.30% and EER is 26.42%. Using L2 Mandarin as the training and test set shows that the tone recognition error rates is 36.27% and EER is 30.19%. Final, we combine the adapted corpus then the recognition error rate is 42.16% and EER is 28.70%. According to experimental results, the recognition results of Mandarin L2 recognizer is better than TCC300 recognizer. The main reasons that speak Mandarin tone are difference between German and Taiwan, and the German speak speed also slower than Taiwan. In the future, we can consider the speed of speak to research and improvement.

並列關鍵字

CALL Tone Recognition Error Detection

參考文獻


[8] 官英華和廖憲正,自動化華語聲調錯誤偵測系統之初探,中原華語文學報 4,pages 117-137,2009。
[3] Y. Qian, T. Lee, and F. K. Soong, “Tone recognition in continuous cantonese speech using supratone models.”, Journal of the Acoustic Society of America, 121(5):2936–2945, May 2007.
[4] C.-Y. Chiang, X.-D. Wang, Y.-F. Liao, Y.-R. Wang, S.-H. Chen, and K. Hirose, “Latent prosody model of continuous mandarin speech.”, In Proc. of the ICASSP, volume 4, pages IV–625 – IV–628, Honolulu, Hawaii, USA, April 2007.
[6] 陳江村,華語發音評量與聲調辨識研究,博士論文,國立清華大學資訊工程學系,新竹,2008。
[12] ACTFL, ACTFL Chinese Proficiency Guidelines. Foreign Language Annals, 20, 471-487, l987.

延伸閱讀