本研究主要針對語者確認系統提出新的辨識流程，係利用動態微分核函數結合高斯混合模型及支撐向量機模型，藉以提升系統效能。 此系統主要是將原始語音特徵向量及已經由UBM-MAP所調適出的高斯混合語者模型參數透過微分核函數中的Fisher Kernel及Likelihood Ratio Kernel做映射，藉此得到不同語者新的特徵參數，藉以建出一個超級向量，在訓練階段中，需將超級向量做正規化，之後利用正規化後的超級向量訓練SVM模型，而在仿冒語音的選取上，則是選取與目標語者模型分數最高的前20名仿冒語音，使得訓練出來的SVM模型更有鑑別力，而測試時，則將測試語音做映射，所得到的超級向量在正規化後與對應的SVM模型計算距離值。 從實驗結果顯示，高斯混合模型選定為128-mixture且選取20位仿冒語者的語音，系統可達最好的相等錯誤率及決策成本函數分別為8.69%及0.1023，比起使用Fisher Kernel做映射的語者確認系統的效能11.92%及0.1500改善了3.23%及0.0477，而比起傳統語者確認模型的效能15.87%及0.1911，改善了7.18%及0.0888。 This thesis proposes a new recognition system to improve performance for speaker verification. The proposed system combines the Gaussian Mixture Model (GMM) and Support Vector Machine (SVM) with Derivative Kernel. Target GMM for each target speaker was adapted from UBM by using target speech. we take Imposter speech feature selection method to choose specific imposter speech features and we can get new features in the target speaker GMM space by Fisher and likelihood ratio kernel mapping. Then we used the new features to establish target supervector and imposter supervector. In the train stage, we used the target supervector and imposter supervector to train SVM model. In the testing stage, we take the supervector into SVM to calculate the distance. The experimental results shows that with a 128-mixture GMM and choose twenty imposter speech, the proposed system obtains a 3.23% EER and 4.77% DCF improvement over the SVM-fisher speaker verification system, and a 7.18% EER and 8.88% DCF improvement over the baseline system.
