在本論文中，我們主要是利用最小錯誤鑑別式(Minimum Classification Error, MCE)重新訓練語者模型，而使用最小錯誤鑑別式(MCE)在訓練語者模型時，所會遇到的最大問題則是要以何種標準選取競爭語者群，針對這一項問題，我們共提出四種競爭語者群的選取方法，包含：排名法、臨界值法、分數分類法及模型分類法，分數分類法及模型分類法皆是將語者參數輸入至支撐向量機(SVM)內做分類的動作，分數分類法是輸入每一位語者的最大相似分數，而模型分類法則是輸入每位語者的模型參數。將參數皆輸入至支撐向量機(SVM)後，再藉由支撐向量機(SVM)優良的分類特性，從語料庫中找到更合適的競爭語者群，進而提升系統語者辨識率，分數分類法對傳統高斯混合模型(Gaussian mixture model, GMM)語者辨識系統有42.27%的錯誤改善率，本論文實驗中是使用TIMIT語料庫為基礎。 In this thesis, we re-train speaker model by Minimum Classification Error Method (MCE). For Minimum Classification Error Method, searching competitive speakers is the most important problem, and then we propose four methods for searching competitive speakers, ex: ranking method, threshold method, model classification method and score classification method. For model classification method and score classification method, we use speaker’s parameters as inputs to train Support Vector Machine (SVM), and SVM will classify target speaker and competitive speakers. In this paper, we expect that the two methods will raise speaker recognition rate. The experimental result shows that Score classification method obtains a 42.27% speaker recognition rate improvement over Gaussian mixture model (GMM). This paper is based on TIMIT database..
