獨立元件分析於單一混音頻道之辨識及分離研究

未知訊號分離(Blind Source Separation)是去從所觀測到由原始信號所產生的混合中找出最根本的因子(Factors)或元件(Components)。獨立元件分析(Independent Component Analysis)是一個用來達成未知訊號分離的一個統計和計算的技術。大部分的獨立元件分析技術都是著重在混合信號的數目大於或等於原始信號的數目。我們特別著重在從一個單一混合信號中來做兩個原始信號的分離。本論文將介紹獨立元件分析的相關背景知識。包含獨立元件分析的一些基本定義，獨立元件分析的前置步驟可以加快其收斂的速度，獨立元件分析依據非高斯的最大化(ICA by Maximization of Non-Gaussianity)和獨立元件分析依據最大似真估計(ICA by Maximum Likelihood Estimation)。此外，頻率域的獨立元件分析也一併被描述。接著，我們將詳細介紹單一混音頻道之聲音辨認及分離的演算法。它用一般化高斯獨立元件分析(Generalized Gaussian ICA)來找出在訓練集合聲音源中的一群事前時間域的基底函數(Time-Domain Basis Functions)。聲音的辨認是依據代數的矩陣距離索引(Algebraic Matrix-Distance Index)來完成；而聲音的分離是依據最大似真估計方法完成。我們也呈現聲音辨認及分離結果的實驗。最後，我們將提出未來的研究方向與改進方式。

關鍵字

並列摘要

Blind source separation is to find the underlying factors or components from the observed mixtures of unknown source signals. Independent Component Analysis(ICA), which is a statistical and computational technique, is used to fulfill blind source separation. Most ICA techniques are implemented on that the number of mixtures is equal to or greater than the number of sources. In particular, we focus on separating two unknown sources from one mixture. This thesis will introduce the background knowledge of ICA. It contains the introduction of some basic de¯nitions of ICA, the preprocessing technique needed for the acceleration of convergence, ICA by maximization of non-Gaussianity and ICA by maximum likelihood estimation. Besides, ICA in the frequency domain will also be described. Then, we will describe the algorithms for single channel source identification and separation. It exploits a priori sets of time-domain basis functions learned by generalized Gaussian ICA for the sound sources given in the training set. The source identification is based on the Algebraic Matrix-Distance Index (AMDI) algorithm, and source separation is based on the maximum likelihood estimation. Some experiments that identify the two musical instruments involved in a symphony and separate the sound coming from each of them are demonstrated to investigate the performance of the approach. Finally, we will address our future works and give some directions on future researches.

並列關鍵字

ICA

參考文獻

[1] F. Asano, S. Ikeda, M. Ogawa, H. Asoh and N. Kitawaki, "Combined Approach of Array Processing and Independent Component Analysis for Blind Separation of Acoustic Signals," IEEE Trans. on Speech and Audio Processing, vol. 11, no. 3, pp. 204-215, May 2003

[2] A. K. Barros and A. C., "A fixed-point algorithm for independent component analysis which uses a prior information," 5th Brazilian Symposium on Neural Networks, pp. 39-42, December 1998

[3] M. S. Bartlett, J. R. Movellan and T. J. Sejnowski, "Face Recognition by Independent Component Analysis," IEEE Trans. on Neural Networks, vol. 13, no. 6, pp. 1450-1464, November 2002

[4] A. Bell and T. Dejnowski, "An information-maximization approach to blind separation and blind deconvolution," Neural computation, vol. 7, pp. 1129-1159, November 1995

[5] A. HyvÄarinen, J. Karhunen and E. Oja, "Independent Component analysis," Wiley, 2001

主題瀏覽