利用次頻域之語音濾波

Abstract

在我們的生活環境中, 隨時充斥著各種聲音. 當我們與人交談時, 這些聲 音便會夾雜於我們的語音之中成為雜訊. 若這些聲音過大, 交談的雙方便 無法確切的知道對方的意思而造成溝通困難! 因此, 在某些特定的環境 中, 我們必須設計濾波器 (filter) 來濾除這些雜訊. 傳統的方法是將語 音訊號用一個高階的自動回歸程序(AR process)來模擬, 然後利用卡門濾 波器 (Kalman Filter) 來濾除雜訊.而高階的卡門濾波器因包含很多的矩 陣運算, 因此計算量相當大而且不容易做成硬體.在這篇論文中, 我們利 用次頻域頻譜較全頻域頻譜來得平緩的特性, 在次頻域的每個頻域分別利 用低階的自動回歸程序來模擬次頻域的語音訊號,於是可以在每個次頻域 使用較低階的卡門濾波器來濾除雜訊. 如此不但可以節省計算量, 而且較 容易做成硬體.實驗結果證實, 利用次頻域的語音濾波的確可以得到很好 的結果. The main goal of speech enhancement is to filter speech contaminated by additive white or colored noise. Conventional apporach is to model speech as high order time varying AR process and apply the Kalman filter. However, the conventional apporach has two problems. First, the Kalman filter, which involves intensive matrix computation, is diffiuclt to be implemented. Second, the estimation of AR parameters from noisy speech is not trivial. In this thesis, we propose new apporaches to solve these problems. We model speech as AR processes with its subband signals and then apply Kalman filtering algorithm to filter noise. Since the lower order AR models can be used for the subband signals, we can apply a lower order Kalman filter, which does not require intensive matrix operation and can be easily implemented. To solve the second problem, we propose adaptive algorithms for white or colored noise respectively. Simulation results show that the speech enhancement in subband domain has much better performance than that in fullband domain.