傳統Fast ICA的獨立成份分析(Independent Component Analysis, ICA)方法會面臨兩種缺點，一個是獨立成份的順序無法決定，另一個是Fast ICA也僅限於解決靜態資料。為了要避免這樣的問題，本研究提出一種基於Robust ICA與頻域訊號分離改進的方法。Robust ICA是一種利用四次多項式的根來尋找分離矩陣，達到收斂效果以求得獨立成份，提供了複數運算的支援並減少白化的過程。將ICA的操作過程放到頻域中執行可以將龐大的問題分成好幾個子問題個別處理，利用硬體平行處理的特性同時分離好幾個聲音段可以加快語音訊號分離處理速度，另外一個好處是在頻域中非高斯的情況會比時域中更加明顯。實驗結果可以發現，基於此架構設計出的語音訊號分離硬體加速器可以解決語音訊號順序不確定的問題，達到語音訊號批次處理的效果。在語音訊號執行速度上，語音訊號分離的處理速度比軟體快，並可以解決實驗中無法達到的語音即時處理的效果。;The traditional independent component analysis of FastICA faces two of disadvantages, undetermined component of order and offline experiment. We proposed a new solution by using RobustICA algorithem with demixing signal in the frequency domain. RobustICA separate independent component by searching demixing vector and using four degree polynomial, supporting complex calculation without whitening progress. The progress which separated component in frequency domain will speed up separated by divided a bunch data into frequency bin with hardware characteristic of parallel process. Moreover, the Non-Gaussianity is obviously in frequency domain. Its shows in experiment that the voice signal separated accelerator have many characteristics including batching progress, ordering component and real time signal separated process.
