獨立成分分析結合子空間增強於時域音訊分離之分析探討

Abstract

音訊分離一直是訊號處理上想要達成的目標，若能從眾多音訊中擷取出自己需要的訊號，在後續也有相當廣泛的應用，本文利用獨立成分分析法，對未知聲源的訊號進行訊號資料分析，以解決盲訊號分離問題。 獨立成分分析法係假設多支麥克風的量測訊號是由多個不同聲源訊號受空間因素影響後混合而成。故本文利用快速獨立成分分析法，針對以兩支麥克接收兩個聲源訊號的情況進行訊號解混合之演算，輸出接近原本各聲源訊號的結果。 在現實情況中，時域訊號會因聲源到麥克風的距離差產生時間延遲的情形，導致分離結果不佳。為了避免此問題，以往的研究大多利用傅立葉轉換將訊號轉至頻率域後再對各頻帶訊號進行分離，但是獨立成分分析法存在不確定性，會導致在將各頻帶分離後訊號加總還原時產生混淆，故需搭配其他理論進行更大量的計算使分離結果更準確。 本文的重點為增加演算法之前處理和後處理步驟，配合獨立成分分析法的特性對輸入訊號先進行時間位移處理，即可改善分離效果，並利用子空間語音增強法對分離後的訊號進行優化。除此之外，本文也對不同聲源組合進行實驗，並對分離結果作分析與比較，最後將整個演算流程以程式實現。

Audio separation has always been the goal of signal processing. If we can extract the signals we need from many audio sources, and have a wide range of applications in the future, this paper uses independent component analysis to signal the signals of unknown sound sources. Analysis to solve the problem of blind signal separation. Independent Component Analysis(ICA) is the common algorithm to solve Blind Source Separation(BSS) problem. By using iteration algorithm, ICA can estimate the most optical demixing matrix for mixed signal. Theoretically, ICA can separate each voice which is made by different source from measured signal which are mixed. However, using time domain ICA algorithm will cause time delay difference problem because the distance between the signal source and each sensor is different. Even though we can transform measured signals into frequency-domain by Fourier Transform and avoid the problem, the ambiguities of ICA will cause dilation problem and permutation problem. The topic of paper is adding pre-processing step for solving and time delay difference problem. In addition, we use subspace speech enhance as post-processing to optimize ICA result.