盲訊號分離的研究可以分為即時混合(Instantaneous Mixture)和旋積混合(Convolutive Mixtures)，在即時混合的研究上，已有不錯的成果，因為解混合矩陣在訓練係數過程中需要很大的運算量，所以將這個部分實現在超大型積體電路(Very-large-scale integration , VLSI)是一個不錯的選擇，本論文想在解旋積混合的硬體做一些改進，希望可以加快處理速度，應付未來應用上之需要。我們主要採取的演算法有兩種並將其實現為超大型積體電路，其中第一種為Infomax演算法，我們使用Torkkola所提出的架構來實現旋積盲訊號源分離，Torkkola的學習規則近似於最小均方誤差演算法，所以我們利用近似於最小均方誤差適應性濾波器的延遲最小均方誤差適應性濾波器來做修改應用到旋積盲訊號源分離中；而第二種為時頻聚類之盲訊號源分離，主要是將訊號做特徵擷取後再經過聚類演算法來達到訊號分離的效果，在這我們提出了特徵節取和K-means的硬體架構。此外我們利用了壓縮感測來增強和重建其分離的訊號，最後也提出了壓縮感測中Orthogonal Matching Pursuit的硬體架構。Blind source separation (BSS) of independent sources from their convolutive mixtures is a problem in many real world applications. In this paper, we design two VLSI architectures for convolutive BSS (CBSS). The first is based on Infomax algorithm and the BSS structure proposed by Torkkola is utilized. As its learning rule is similar to least mean squares (LMS), we apply delayed LMS (DLMS) to BSS. The proposed architecture based on sharing multiplication improves adaptation delays and critical path. The second VLSI architecture is based on time-frequency masking based BSS. This method generates useful features for each time-frequency points and then clusters them to achieve signal separation. For this algorithm, we propose VLSI modules for feature generation and K-means and orthogonal matching pursuit.
