The Automatic Design of the Decision Tree Recognizer for Small Chinese Vocabulary

本論文主要在探討如何建立一個有效率的相連音國語詞彙辨識系統之參考 樣本空間，我們利用中文語音的特性，採單音節的辨識方式來做詞彙的辨 識，同時我們也利用欲辨識詞彙集合中，每一單音節間相似程度不同的差 異，發展出音節測距函數，根據此函數的測距我們將參考樣本空間做適當 的分類，並將決策樹的樹狀結構運用在分類上，因此我們也設計出可以依 據欲辨識詞彙集合組成的不同而能自動產生一棵有效率的辨識決策樹的方 法。實驗分成語者獨立和語者相關兩部份，辨識詞彙集合則以台灣區26個 縣市地名及常用的30個中文名詞為測試語料，在語者獨立實驗上我們達 到90﹪的辨識率，在語者相關實驗上我們則達到95﹪的辨識率。The major goal of this thesis is to discuss how to build an efficient recognition reference space for the small Chinese vocabulary set. On the basis of similarity of Chinese isolated syllable, we develop a "distance function" by which we can properly classify the reference space. As a result, we design a method that can automatically generate decision tree recognizer for each vocabulary set. There are two parts in our experiments -- "speaker dependent" and "speaker independent." Our testing vocabulary sets include twenty-six city names of Taiwan and thirsty Chinese vocabularies. The experimental results showed that we achieved recognition rate 90﹪in "speaker independent" and 95﹪in "speaker dependent"