台語語音辨識及智慧型口語對話汽車導航系統

Abstract

在本論文中建立了一個台語音節辨認器並且將之應用於一個實際應用系統-「慧型口語對話汽車導航系統」中。在論文中首先針對台語語音特性入聲調的部份深入研究，我們由confusion matrix去分析入聲調對辨識系統的影響性，且根據語言學上的知識修改文字拼音資料庫，使得訓練出來的聲學模型更為可靠，進而提升辨識率；並利用Kullback Leibler(KL) distance去觀察聲學模型在修改前與修改後之間的差異。同時論文中也將sub-syllable bigram 語言模型(Language Model)加入辨識系統中，使的辨識率得以提升。在論文最後把台語語音辨識系統應用在-「慧型口語對話汽車導航系統」。

In the thesis, a syllable recognition system for Taiwanese was established and applied to a real application - Intelligent Transportation System(ITS). First , a syllable-based Taiwanese speech recognition system was implemented. And the effect on the recognition performance of the syllable with entering tone was carefully exiamined. Based on the linguistics knowledge, the syllable with entering tone in database was re-labeled according to the tone sandhi of Taiwanese in order to improve the recognition rate. In addition, Kullback Leibler distance between the acoustic model before and after re-labeling was examined to verify the linguistics knowledge. And the syllable bigram language model to the recognition system will obtain the rate improvement. Finally, the Taiwanese Speech Recognition System was applied to an Intelligent Transportation System(ITS)