盲訊號源分離在各個領域如今是一個很重要的研究方向,例如:生物訊號處理,腦科學,多媒體訊號處理…等,在語音訊號處理更是一個很好的研究主題。而在解盲訊號源分離這個問題上,是需要很大的運算量,所以在未來如果要即時應用下,實現成超大型積體電路是一個很好的選擇。本論文主要採用的演算法是時頻聚類之盲訊號源分離演算法,主要將訊號的特徵擷取出來後,利用聚類演算法將訊號分離,來得到我們想要的訊號。在這邊我們利用壓縮感測重建並加強分離之訊號,並且在壓縮感測中提出一個在面積上較有效率的Orthogonal Matching Pursuit硬體架構。而在壓縮感測演算法中,我們必須利用一個訓練好的字典來完成其重建訊號的動作,在這邊我們是利用K-SVD演算法來訓練字典。此演算法在各個領域也是一個重點的研究方向,例如:影像處理,訊號去噪…等,而這個演算法運算量一樣龐大,若我們想即時的運用,實現成超大型積體電路勢在必行。所以在這邊我們也提出了一個K-SVD的硬體架構,利用此硬體架構,可以有效的減少利用K-SVD所花費的時間。 Now, blind source separation (BSS) is a very important research direction in various fields, For example, Biomedical Signal Progressing, brain science, multimedia signal processing, etc. And in audio domain, it is a good theme, too. In the paper, we design a BSS VLSI architecture which based on time-frequency masking. We fetch the feature for each time-frequency points and cluster them to get the separation signal. In the architecture, we add the sharing multiplication, and modify it to achieve more area- efficiency. On the other hand, when we use the algorithm to solve the problem, we must use a trained dictionary, and the training time is so long. To solve the question, we also design a VLSI architecture based on K-SVD algorithm. We use the architecture of orthogonal matching pursuit we designing to implement the K-SVD, and we also design an Acceleration & Lock Unit to decrease the power and the clock rate.