動態臨界值資料選擇仿射投影演算法運用於聲學回聲消除

Abstract

本論文針對線上會議與遠距通訊等經由網際網路的語音傳輸中經常發生的聲學回聲干擾問題，提出結合仿射投影演算法（Affine Projection, AP）與資料選擇法（Data Selective, DS）的動態臨界值資料選擇仿射投影演算法（Dynamic Thresholds Data Selective Affine Projection Adaptive, DT-DSAP）用以達成聲學回聲消除（Acoustic Echo Cancellation, AEC）工作。我們對 DS 的各臨界值作出嚴謹的討論，在誤差訊號、期望訊號以及輸入訊號方面，皆得出足以區分主要更新訊息、不具創新度訊息及異常值干擾的動態臨界值，幫助了演算法在系統處於追蹤時期使用較多的運算量發揮 AP 的快速追蹤特性，在穩態時期則節省運算量與增強強健性能，並且更有效的對語音訊號等具過小訊息量時間段特徵的輸入訊號進行處理。另一方面，我們以不獲取非更新訊息的方式將 DS 與 AP 結合，可解決既有結合方式可能造成的幾項重要問題，使演算法理想地發揮 DS 的效果。模擬結果分析討論了 DT-DSAP 在 AEC 應用中各臨界值的參數選擇與表現，結果顯示所設置臨界值可如預期的發揮各自功效。在與既有演算法的比較中，我們使用多種輸入訊號在不同強度的異常干擾情境進行 AEC 應用模擬，DT-DSAP 在收斂及追蹤速度、強健性能與運算量的節省展現出更佳的平衡與優勢。

In this paper, we propose Dynamic Thresholds Data Selective Affine Projection (DT-DSAP), which combines Affine Projection (AP) and Data Selective (DS), to address the acoustic echo cancellation (AEC) problem often faced in voice transmis sion over the Internet. Our proposed DS method yields dynamic thresholds for error signals, desired signals, and input signals that are sufficient to distinguish between major updates, non-innovative signals, and abnormal interference. This helps the adaptive algorithm to utilize the fast tracking feature of the AP with more compu tation when the system is in tracking period. During steady state, the algorithm saves computation and enhances robust performance. In addition, the algorithm can efficiently process the input signals with the characteristics of the excessively small message volume time periods, such as voice signals. On the other hand, we propose to combine DS with AP in a way that does not acquire non-updated infor mation, which can solve several problems that would be caused by the established combination approach, and enable the algorithm to optimally utilize the effect of DS. In the simulation results, we examine the performance and parameter selection of DT-DSAP in AEC applications, and the results show that each threshold can perform as expected. In comparison with existing algorithms, we simulate AEC applications using multiple input signals in abnormal interference scenarios of different intensities. DT-DSAP shows a better balance of tracking speed, robustness, and computational savings.