應用高斯混合模型局部調整學習於語者辨識之研究

Abstract

本論文主要在探討高斯混合模型(Gaussian Mixture Model, GMM)的學習與 其在語者識別(Speaker Identification)上的應用. 在先前的研究中,用語者的GMM來做語者識別已經有很不錯的成果, 對GMM的高斯元件個數也有探討. 在本論文中我們討論到SLUG(Supervised Learning and Unsupervised Growing)訓練模型的想法. 改進為單群的調整,並且以TCC-300麥克風語音資料庫作為我們實驗的語料, 以實驗證明其可行性.

This thesis discuss the Gaussian Mixture Models learning and use it on speaker identification. In pass research,it has the good result with using GMM on speaker identification,and someone maybe debate with the initial numbers of GMM components. In this thesis,we take one thinking SLUG(Supervised Learning and Unsupervised Growing) to train our GMM models,and imporve our tuning step to just tuning a single cluster. finally,we use the TCC-300 microphone speech database to be our experiment data to prove our theory.