應用頻域獨立成份分析法及頻率成份調整分離鳥聲研究

本篇論文採用頻域獨立成份分析法(Independent Component Analysis, ICA)分離混合鳥聲訊號，並將分離之訊號以頻率成份調整技術(Frequency Component Modification Techniques, FCMT)調整權重值提高分離鳥聲之清晰度，並以動態時間校準法(Dynamic Time Warping, DTW)表示分離之程度好壞。獨立成份分析法採用現今較熱門的快速獨立成份分析法(Fast Independent Component Analysis, ICA) 分離，利用其較簡單之演算法縮短分離時間，若把戶外實際鳥聲鳴叫時經過樹木、建築物之空間折射、散射等空間效應都納入將會使獨立成份分離法變成複雜的運算，在時域上難以實現，在此利用頻域獨立成份分析法能夠將空間效應等在數學式中使用複雜之摺積運算轉換為簡單的相乘動作並能夠完整呈現真實情況下的分離情形，另採用調整頻率成份調整頻率臨界值以獲得更為乾淨的鳥聲，實驗中採用動態時間校準法比對分離後的鳥聲與原始鳥聲是否相似?實驗數據證明若加以頻率成份調整技術分離之鳥聲會更接近原始鳥聲獲得較佳的分離效果。

關鍵字

獨立成份分析法；頻域成份調整；動態時間校準法

並列摘要

The thesis applies the Independent Component Analysis(ICA) on the frequency domain to separate the mixed bird sounds and uses the Frequency Component Modification Techniques(FCMT) to adjust weight value to enhance the cleanness of separated bird sound. The quality of the separation can be shown by the Dynamic Time Warping(DTW). The Fast Independent Component Analysis (FastICA) is a popular ICA algorithm to separate signal. The FastICA is known of having Fast computation and simple syntax. If the spatial refractions and scattering of trees and buildings were included, the operation of ICA will be more complicated. The spatial effect is very difficult to realize in the time domain; therefore we use ICA on the frequency domain to solve the issues. Hence, the convolution operation can be replaced by multiplication operation. The FCMT can adjust the weights of frequency threshold to obtain a clean bird sound. Finally, the DTW can be used to measure how close the separated bird to the original bird sound. According to the results of the simulation, the ICA included FCMT provides better performance than using the FastICA alone on the frequency domain.