Abstract
在本篇論文中，我們嘗試用隱藏式馬可夫模型來處理台語的大辭彙語音辨識。 我們所採用的基本語音單位是右連音，若只考慮音節內的關連，辨識結果達到88%。若更進一步考慮音節之間的關連，則辨識結果可達到92.11%，但同時也使模型總數遽增四倍，為了解決模型過多但訓練語料不足造成的問題，我們對模型內的狀態進行分類，採用的方法是語音分類決策樹，狀態總數降為原來的3/5，且辨識率也增加到92.87%。In this thesis, we use HHM to deal with the problem of largevocabulary recognition task of Taiwanese. The phone units weuse are right context dependent (RCD) phonemes. If we justconsider the effect of inside syllable context dependency, thetop 1 recognition accuracy reaches 88%. If we modify our modelsto include all inter-syllable context dependency, the result oftop 1 recognition accuracy increases to 92.11% but the totalnumber of states in our network increases to 4 times. Tocompensate the insufficient training data due to the largeamount of models, we use the method of Acoustic Decision Tree todo state clustering.The total number of states decreases to 3/5,and the recognition rate also increases to 92.87%.