音響回音消除與噪音控制

Abstract

在免持聽筒的語音通訊系統中，藉由回音與噪音的消除，來提升語音的品質是很重要的。本論文將探討回音與噪音控制的相關問題與技術。 在回音消除方面, 傳統的MLC (maximum length correlation) 法估計房間脈衝響應時．深受遠端和近端語音的影響; 尤其在 “雙邊對話” (double-talk)時，濾波器會做錯誤的調整而失效。 我們提出Iterative MLC (IMLC) 法，排除遠端語音對濾波器係數的影響。它在單邊遠端語音發話時，效能和一般的LMS適應性濾波器一樣好，同時運算簡單。由於近端語音的影響還不能完全排除，我們再提出IMLC AEC and Double Talk Detection (IMLC&DTD)架構加以改進。這個架構主要的概念是觀察濾波器係數的平方值(square value) 變化的情形，我們分別假設了它們的機率分佈，進一步發展一個Maximum Likelihood Test 偵測器並做統計分析之Receiver Operating Characteristic (ROC)曲線，以尋求較佳的Neyman-Pearson decision rule作為偵測器的法則。它能有效區別 “雙邊對話”，或 “脈衝響應在變化” 兩種狀況。我們也從電腦模擬分析得到證明。 在噪音控制方面，我們針對密閉管路中的主動式噪音防治提出 “一維映像法” 的理想研究模型，這個一維模型能用來模擬管路中聲源與各感測器之間的脈衝響應。獲得這些脈衝響應之後就可以用 Least-squares 方法求得所需的濾波器。目前我們用電腦模擬發現只用兩個感測器置於管路的兩端即能良好估計管路中聲音的能量。我們也找出(1).濾波器係數與管路中聲源位置之間的Coefficient invariance關係(2).聲源位置與噪音最低能量的reduced energy invariance關係，(3).觀察出設計該濾波器的簡單方法。我們完成所有演算法之電腦模擬分析並得到證明。

In hands-free communication systems, the enhancement of speech signals by canceling the acoustic echo and the disturbing ambient noise is particularly important. In this thesis, the fundamental problems and techniques of echo and noise cancellation are introduced. The conventional maximal-length-correlation (MLC) algorithm to estimate room impulse response for acoustic echo cancellation (AEC) is disturbed by both far-end and near-end speeches. We propose a new Iterative-Maximal-Length-Correlation (IMLC) algorithm to reduce the far-end speech interference. To avoid the near-end interference, a new double-talk detection method is proposed by tracking the squared coefficients errors of the AEC filter. Based on the assumption of uniformly distributed time-variant echo-path-change, we develop a simplified likelihood ratio test and plot the detection performance by a Receiver Operating Characteristic. Computer simulation shows the proposed DTD method is effective in discriminating double-talk from time-variant echo-path-change. Active noise control (ANC) involves the determination of the impulse responses between the sources and sensors; however, these responses are affected by the number and locations of sources and sensors, reflection coefficients, etc. and the influence of these factors on the ANC performance remains unclear. We first modify the commonly used three-dimensional image method to a one-dimensional case and use it to simulate impulse responses in a duct so that an optimal ANC filter can be easily solved by the least-squares method. Based on the one-dimensional image model and the orthogonal property of the impulse response matrix, the filter impulse response has been shown to have periodic pairs of non-zero coefficients. We find that the magnitudes of the coefficients and the reduced ANC energy are fixed when the primary source location is changed. The coefficient invariance property also agrees with the absorbing termination and is confirmed by simulations.