臺灣博碩士論文加值系統

　　本論文中，探討特定語者（Speaker-Dependent）之中文語詞辨識（Mandarin Phrase Recognition）系統如何實現於中量字彙（Medium-Vocabulary）的連續語音辨識上。由於，自然流利的連續語音會產生連音效應（Coarticulation effects），造成連續語音辨識的困難。所以，本文首次嘗試以語詞為辨識的基本單位，可以避免連音效應，同時，簡化辨識的處理過程。
　　本系統以卡氏轉換（Karhunen-Loeve Transform）作為語音頻譜特徵之萃取方法，採用機率統計法則的二次式分類器（Quadratic Classifier）作為後級辨識器。由於卡氏轉換具有最小截斷誤差的特性，於是，語音頻譜資料經卡氏轉換後，可以將能量集中在前幾個維度，可以節省系統的計算資源與運算的時間。因為二次式分類器具有最小分類誤差的特性，所以，可以利用機就緒統計的模式來計算每個語詞的變異性，降低語詞變異性對辨識率的影響。
　　由實驗結果得知，語詞數目200個，辨識率為93.5％。相較以歐氏距離為後級辨識器的辨識效果提升了約30％。

　　This paper is proposed to realize on medirum vocabulry of speech recognition. Because there is difficulty, such as coarticulation, on continuous speech recognition. We attempt recognition with phrase for the first time in order to avoid the difficulty and reduce complex process.
　　This system is based on Karhunen-Loeve Transform (KLT) for extracting spectral feature and Quadratic Classifier. As the property of minimum mean square truncation error for KLT, only the first components associated with the largest eigenvalues are required to preserve. most of information if the phrase data set. The property of minimum classification error for Quadratic classifier is employed that characterize the variability in the phrase.
　　The proposed phrase recognition system differs from existing Euclidean Distance measure. A database, with 200 Mandarin phrase,m is collected for system evaluation. It is demonstrated that over 93.5% correct classification rates can be achieved.