多通道語音強化使用相對轉移函數建構之零波束形成

Abstract

本論文提出一個針對穩態或是非穩態干擾聲源消除的語音強化方法。消除非穩態雜訊是目前語音純化研究中相當重要的問題，本論文提出一個以適應性濾波器為基礎並結合零波束形成演算法的空間濾波器。零波束形成演算法是利用奇異值分解法找出干擾聲源的零空間當作零波束形成器。論文中零波束形成器分為固定式和可變式。可變式零波束形成器以階數回歸最小平方誤差估計的方法找出當前麥克風收到訊號的子空間，並使用子空間相似度的演算法，剔除目標聲源子空間並利用正交上三角分解產生一組獨立基底。這些基底組成了干擾聲源的子空間。零波束形成演算法可應用在不同適應性濾波器上，本論文將固定式零波束形成器應用在廣義旁瓣對消器和參考訊號架構為基礎之濾波器；可變式零波束形成器則應用在廣義旁瓣對消器。所提出的零波束形成演算法同時可對目標聲源做語音活動偵測以加強適應性濾波器的效能。本論文最後以線型麥克風陣列在實際環境下的實驗結果說明本演算法的效能。

The thesis proposed a speech enhancement method for stationary and nonstationary interfering sources. To effectively eliminate nonstationary intereferences is an important research topic for speech enhancement. This thesis proposed an adaptive nullforming spatial filter. The nullforming algorithm uses singular value decomposition (SVD) to find the null space of interfering sources. Both fixed and adaptive nullforming algorithms are studied. The adaptive nullforming uses order recursive least square estimation (ORLS) to find the subspace of presently received signal. The algorithm assumes that the relative transfer functions (RTFs) of sources from different direction can be obtained. The estimated subspaces from these RTF’s contain the subspace of the desired signal. They are sorted according to the distance to the subspace of source from every direction. Then the bases of desired signal subspace from estimated subspace could be removed and a set of independent basis are derived using the orthogonal triangular decomposition (QRD). The basis then comprises of the subspaces of the interfering sources. The fixed nullforming algorithm could be appiled to generalized sidelobe canceler (GSC) and reference signal based adaptive beamformer (RSAB) while the adaptive one can be applied to GSC. Further, it can also be used as directional voice activity detection (VAD) to enhance the performance. Finally, experiments using a linear microphone array under real environment are conducted to demonstrate the performance of proposed algorithm.