基於獨立成分分析(ICA)與廣義機率遞減法則(GPD)之語者辨識與確認技術

Abstract

本論文提出一個新的詞語不相關的語者辨識系統。使用常見的獨立成分分析法（Independent Component Analysis, ICA），找出原始特徵－梅爾倒頻譜參數（Mel-Frequency Cepstral Coefficient, MFCC）中蘊含重要資訊且互相獨立的成分，並將這些獨立成分用於特徵轉換上。此外，運用找出來的ICA基底進行降維的動作。因此，在所提出的辨識系統中，以ICA基底所轉換出來的特徵做為替代MFCC的新特徵。由實驗結果可以證明，使用新特徵的辨識結果比MFCC的辨識結果佳。而分類器方面，應用廣義機率遞減法則（General Probability Descent, GPD）對高斯混合模型辨識器（Gaussian Mixture Model, GMM）做最佳化的動作，以取代傳統上使用的最大相似度法則（Maximization Likelihood, ML）。由於GPD的目標是直接對辨識錯誤率做最小化的動作，因此GPD將決策規則（decision rule）以函數的形態納入整體架構中，故GPD適合用來最佳化辨識模型的參數。在以GMM為主體的系統中，本論文將具體實現GPD法則。經實驗證明，和傳統以MFCC及ML做為最佳化GMM的架構相比，本論文所提出的新架構有較佳的辨識結果。

In this thesis, we propose a novel text-independent speaker recognition system. A decomposition called the independent component analysis (ICA) is used to find out the most important and independent components of the original feature MFCC for the process of feature transformation. We also can reduce the dimension of the features depending on the ICA basis. These ICA-based features are used as our new features in the proposed system. The experiments have shown that using new features has an improvement on using MFCC. In addition, in the classifier phase, we apply the general probability descent (GPD) method to optimize the GMM recognizer instead of the conventional method such as the maximization likelihood (ML) method. That’s because the objective of GPD is to minimize the recognition error rate directly. The decision rule of GPD appears in a function form in the overall criterion and it is suitable for the model parameter optimization. We present an implementation of the GPD method in a GMM-based system. The experiments have shown that the recognition rate of our proposed system is higher than the rate of the system with MFCC as the features and the ML-based GMM as the recognizer. It means that the experimental results verify our proposed system with ICA-based features and the GPD-based GMM recognizer.