應用獨立成份分析法在時域-頻域分離鳥聲之探討

摘要本篇論文研究主要探討如何採用獨立成份分析演算法(Independent Component Analysis, ICA)在時域與頻域上，分離鳥聲效果之差異。在時域上採用快速獨立成份分析演算法(Fast Independent Component Analysis, FastICA)處理。運用FastICA原理，以模擬與實驗方式來針對此演算法之特性進行鳥聲混音分離。在頻域上，先透過傅立葉轉換形式，並進行白化處理與最佳對數頻域預估來進行改善與調整，預估方式可找出適當之頻率語音框並透過這些方法能去除一些雜訊。最後，採用動態時間校準法(Dynamic Time Warping, DTW)將時域與頻域所分離出來之成果與原始鳥聲作比較，依照比較之距離數據來判定，當數據愈小愈接近零時所分離之鳥聲就越接近原始鳥聲，經比較數據結果發現頻域法分離優於時域法分離。

關鍵字

獨立成份分析法；最佳對數頻域預估；動態時間校準法

並列摘要

Abstract This thesis focused on the use of Independent Component Analysis (ICA) algorithm in both time and frequency domain, separates the specific sound of birds from a mixed sound environment. Firstly, in the time domain, one uses the Fast Independent Component Analysis (FastICA) algorithm to simulate how to separate a sound from the mixture sound. In the frequency domain, with the Fourier transforming, whitening and the Optimally Modified Log Spectral Amplitude can improve and allocate dynamic voice frame size in order to remove most of the noise. Finally, the Dynamic Time Warping (DTW) is applied to compare which method is better. The comparison of the distances from the original sound to the separated sound, which are separated by using the time and frequency ICA methods, are being used to show that the ICA on the frequency domain is better than the ICA on the time domain.