自動中文音素錯誤偵測器

Abstract

傳統上音素錯誤偵測器以事後機率作為信心度指數，跟門檻值比較之後可決 定一個音素的正確性，而本論文提出兩種比基本系統好的方法，都是以多層感知 器為基礎的錯誤偵測器，其中最主要的概念是引入多維的事後機率向量，而兩種 多層感知器偵錯系統的差異在於訓練資料的部分，期許利用多層感知器網路的學 習特性，抽取有用的訓練語料，以訓練出效能較佳的多層感知器網路，最後利用 外國人的中文發音語料來測試三個系統的偵錯效能。

Traditionally, phone error detector uses a posterior probability as confidence measure, the correctness of a phone can be decided by comparing it to its corresponding threshold. In this thesis, two systems better than baseline are proposed, both are phone error detectors based on MLP network. The main concept of MLP-based system is the introduction of multiple-dimension a posterior probability. Besides, the difference between the two proposed MLP system is their training data. Hope that we could improve the performance of the MLP network by utilizing its learning property and taking useful data as training data. At last, we test the error detecting performance of these three phone error detectors by the Mandarin corpus of foreign speakers.