適應性波束形成器於寬頻語音純化使用具二階約束之卡曼濾波器

Abstract

當演算法中與環境及麥克風陣列的相關假設是不成立的時候，適應性波束形 成方法的性能會有大幅的衰減。當預設目標訊號出現在訓練的樣本數中，即使在 預設的指向向量和實際的指向向量有輕微的失配發生，陣列演算法的效果也會變得相當靈敏，相關的失配問題發生於變動的環境以及遠近效應、聲源散佈、與局部散射等。 本論文提出一套穩健型寬頻波束形成器，基於最佳化中在最糟情況下解決任 意且受規範的目標訊號之指向向量失配的問題。利用麥克風陣列訊號的空間資訊 ，由最小變異無失真響應(MVDR)的波束形成器以空間濾波的方式針對聲源方向純化語音，同時壓抑來自於其它方向的雜訊。在實際應用的例子，波束形成器可以表示為狀態觀測器於二階展開的卡曼濾波器(SOE-KF)。然而窄頻波束形成器沒有考慮低頻訊號的空間指向性，目標聲源會因此受到壓抑而造成語音的失真。為了在指向向量失配的情況下輸出較高的輸出訊號與干擾加雜聲比，演算法根據不同頻帶下訊號的特性來選擇適當的指向向量的限制範圍。此外，當目標聲源不存在時，演算法對雜訊做追蹤，並透過零波束形成限制式於SOE-KF中以進一步提升語音純化的效果。本論文所提出的方法不僅改善了壓抑雜訊的效果並且提升了語音的品質。透過模擬和實驗驗證，本論文所提出的演算法有效地提升語音品質於吵雜以及有迴響的環境，並與其它已知的方法進行比較和分析。

Adaptive beamforming methods are known to degrade significantly if some of underlying assumptions on the environment, sources, or sensor array are violated. The array performance may become sensitive even for a slight mismatch between the presumed and actual signal steering vectors when the desired signal is present. Such kind of mismatch occurs due to the dynamic environment, near-far mismatch, source spreading, and local scattering.

This paper presents a novel approach to design the robust broadband beamformer against arbitrary steering vector mismatch based on the optimization of worst-case performance. Using the spatial information from the microphone arrays, the desired source is enhanced while suppressing the directive noises via the robust minimum variance distortionless response (MVDR) beamformer. In practice, the beamformer is formulated into state-space observer form of the second-order extended (SOE) Kalman filter. However, the narrowband beamformer won’t consider the signal directivity in the low frequency-bands and the desired source would be distorted. For maintaining higher OutputSINR under steering mismatch, the broadband selection of the steering vector bound is investigated. Furthermore, the noise tracking is utilized as null constraints into the SOE Kalman filter for speech enhancement when the source is absent. The proposed algorithm in this thesis not only improves the performance of noise suppression but also enhances the speech equality. Simulations and experiments demonstrate the effectiveness of the proposed algorithm in a noisy and reverberant environment by comparing with existing algorithms.