摘 要 近些年來，由於語音辨識技術的純熟，語者辨識的研究已經越來越受到重視，且廣泛的應用在各種領域中，而本實驗室對於國語語音辨識的研究已經累積了不少成果，但對於語者辨識這個領域較少涉獵。因此，本論文是針對國語語音資料，建立語者身分識別系統，來對語者辨識做一個初步研究。 在本論文中，以高斯混合模型來代表每一位語者特徵向量的統計分佈，高斯混合模型是語者辨認中最常使用的統計模型，實用上共變異矩陣常假設為對角化型式，這個假設忽略了特徵參數間的相關性，但若使用全共變異矩陣，則計算量及參數量將大大的提昇，而顯的不切實際。因此本篇論文主要是針對幾種不同特徵參數的轉換，將全共變異矩陣轉換為對角共變異矩陣，使系統有最佳效能，並結合向量量化的方法訓練模型，運用於語者識別上。 而實驗中以100位語者，來做語者辨識之實驗，從實驗中可發現經過特徵參數轉換後所求得的對角共變異矩陣，比未做任何處理的高斯混合模型來的好。 Abstract In recent years, the characteristics in human biology, such as the fingerprint, palm prints, eye and voice etc, were used to recognize personal identities. Among these biological characteristics, the human speech has the properties as easiness of product and extraction, as well as transmission through the network of telephone, and therefore suits for application of human identity. The Gaussian mixture model (GMM) is the most using statistical model for speaker recognition. In this plan, use the Gauss mixture model to represent each speaker distribute of feature vector. The covariance matrices can be full or diagonal matrices forms in GMM. Most of these models assume diagonal covariance matrices, but this assumes to ignore the each feature vectors correlation. If we assume the covariance matrix is diagonal form, we must use the orthonormal transformation to derive uncorrelated feature vectors. So that to reduce the relation of each vector of feature to lowest. This technique can improve the performance of speaker or speech recognition system.
