本論文主要針對語者確認系統提出新的確認流程，使得系統效能得到提升。此架構是利用權重幾何組合及權重算術組合，結合傳統的通用背景模型及最佳競爭語者等相似度值計算方法，透過支撐向量機計算出的權重，產生新的目標決策函數，以達到最佳的辨識效果。 此系統主要是將輸入語句的梅爾頻率倒頻譜特徵與註冊語者的高斯混合模型算分數，透過加權幾何組合與加權算術組合，建立出輸入訓練支撐向量機模型及測試的向量。本論文加入了分數尺度化，以調整不合理的分數，尺度化範圍設定在 ，此範圍在實驗中驗證可以使效能再獲得些許的提升。在仿冒語者的選取上，則是選取與目標語者模型分數最高的前60位，使得訓練出來的模型不失鑑別力，又能夠有效的節省運算時間。最後我們將加權幾何組合、加權算術組合與通用背景模型及最佳競爭語者等其他計算相似度比值的方法進行整合，使系統效能再提升。 從實驗結果顯示，在我們使用的架構中，高斯混合模型選定為128-mixture、選取60位仿冒語者及尺度化分數範圍為 ，系統可達到最好的相等錯誤率及決策成本函數分別為6.06%及0.0787，比起參考文獻[50]的相等錯誤率改善了2.63%，決策成本函數改善了0.0236，而比起參考文獻[51]的語者確認系統的相等錯誤率改善了0.70%。This thesis proposes a new verification system to improve the performance for speaker verification. The proposed system combines Weighted Geometric Combination (WGC), Weighted Arithmetic Combination (WAC), Universal Background Model (UBM) and Most Competitive Cohort Model (MAX), and uses Support Vector Machine to generate weight vectors for a new decision function. We calculate the likelihood scores of input utterances’ MFCC with registered speaker model, and build input vectors for training SVM models and testing through WGC and WAC. Also, we proposed a scaling method to adapt the unreasonable likelihood scores. The range of scaling is , and this method and range is shown to improve the system by our experiments. And then we select the Top 60 imposters from total speakers’ likelihood scores by imposter selection. These methods not only can make the training model more robust but also can reduce the time of calculations. The experiments results are based on 128-mixture GMMs, Top 60 imposter selection and scaling range of environment. The proposed system obtains a 2.63% EER and 2.36% DCF improvement over [50], and a 0.61% EER improvement over [51].
