基於傳遞函數比與非線性H∞濾波器之穩健適應性語音純化波束形成器

Abstract

在過去三十年中，利用麥克風陣列純化語音的技巧已受到許多研究者的專注。在許多現實環境中，目標語音訊號通常受穩態雜訊與多個非穩態雜訊所干擾。本論文的目標為利用均勻線性麥克風陣列提供一滿意的波束形成器(亦稱空間濾波器)效能與穩健度對抗背景雜訊與空間響應效應。本論文提出兩種適應性空間濾波器：以傳遞函數比為基礎之適應性空間濾波器與以二階延伸H∞濾波器為基礎之穩健最小變異無失真響應空間濾波器。 在第一類適應性空間濾波器中，傳遞函數比為事先利用系統識別方法來模型。本論文提出的以傳遞函數比為基礎之適應性空間濾波器由傳遞函數比空間濾波器與多通道適應性濾波器所構成。傳遞函數比空間濾波器用以消除多個非穩態訊號中的主要部分，而目標語音訊號的通道效應則由傳遞函數比的資訊來同化。由於H∞ 濾波器能較穩健於模型誤差，因此從傳遞函數比空間濾波器輸出的剩餘雜訊訊號則由限制H∞ 濾波器來消除。此外，本論文提出虛擬聲源的觀念用以簡化多個非穩態訊號的空間複雜度。 在第二類適應性空間濾波器中，傳遞函數假設為一單純延遲模型與一不確定數的組合。本論文提出一新的方法用來實現穩健最小變異無失真響應空間濾波器。穩健最小變異無失真響應空間濾波器是設計在最差效能下最佳化的結果，此濾波器對於目標訊號方向向量誤差提供了絕佳的穩健度。為了方便即時性的實現，此種空間濾波器曾轉化為狀態空間模型並利用二階延伸Kalman濾波器來實現。然而，二階延伸Kalman濾波器假設系統動態為已知並假設雜訊為零平均與已知變異量。此類假設會影響穩健最小變異無失真響應空間濾波器的效能。本論文發展與推導二階延伸 H∞ 濾波器並用以實現穩健最小變異無失真響應空間濾波器。二階延伸 H∞ 濾波器是在最差情況下最小化估測誤差並對雜訊統計特性並無假設。最後，本論文提供模擬與真實環境實驗結果用以驗證演算法效能。

Speech enhancement techniques, utilizing microphone array, have attracted attentions of many researchers for the last thirty years. In many practical environments, the desired speech signal is usually contaminated not only by stationary noise but also multiple nonstationay interferences, such as competing speech signals. The objective of this dissertation is to design robust adaptive beamfromers to reduce background noise and compensate channel effects using a uniform linear microphone array. Two adaptive beamformers, the transfer function ratio (TFR)-based adaptive beamformer and the robust adaptive beamformer based on the second-order extended (SOE) H∞ filter, are proposed in this dissertation. In the first adaptive beamformer, the TFR is obtained using the system identification method in advance. The proposed TFR-based adaptive beamformer consists of the TFR beamformer and multi-channel adaptive filter algorithm. The TFR beamformer is used to block the major component of the multiple interference signals and the associated information is then used to equalize the channel effect of the desired speech. The residual noise from the TFR beamformer output is suppressed by the constrained H∞ filter due to its robustness to the modeling error. In addition, the virtual sound source concept is proposed to simplify the theoretical treatment for multiple competing speech signals. In the second adaptive beamformer, a novel approach to implement the robust minimum variance distortionless response (MVDR) beamformer is proposed where the acoustic transfer function is assumed to be delay-only propagation with uncertainty. The robust MVDR beamformer is to optimize the worst-case performance for an arbitrary but norm-bounded desired signal steering vector mismatch. For real-time consideration, the beamformer was formulated into a state-space observer form and the SOE Kalman filter was derived. However, the SOE Kalman filter assumes an accurate system dynamic and known statistics of the noise signals. These assumptions limit the performance under various uncertainties. This dissertation develops the SOE H∞ filter for the implementation of the robust MVDR beamformer. The estimation criterion in the SOE H∞ filter design is to minimize the worst possible effects of the disturbance signals on the signal estimation errors without a priori knowledge of the disturbance signals statistics. Finally, the results from simulations and practical experiments are provided as proof of the performance of these proposed approaches.