結合適應性波束形成與後濾波進行語音強化

Abstract

本論文提出一套結合適應性空間濾波與後濾波的方法進行語音強化。利用麥克風陣列訊號的空間資訊，以空間濾波方式對聲源方向純化語音，論文中採用成效較佳的Dahl’s濾波器。為了進一步純化語音，我們使用單聲道語音強化的方法進行後濾波。後濾波主要包含雜訊估測與增益函數兩部份，雜訊估測將分別使用長時間語音活動偵測與最小控制遞迴平均法，搭配以頻譜刪減法與對數頻譜幅值計算出之增益函數，並實際比較在高速公路雜訊與音樂雜訊下的純化效果。

An approach combined adaptive spatial filtering and post-filtering for speech enhancement is proposed in this thesis. Using the spatial information of microphone array signals, we purify the speech in the sound source direction by applying spatial filtering. For the spatial filter, we choose Dahl’s beamformer due to it’s relatively better performance. To further purify the speech, we use sigle-channel speech enhancement methods for post-filtering. Post-filtering mainly contains noise estimation and gain function parts. We will use long-term voice activity detection (LTVAD) and minima controlled recursive averaging (MCRA) for noise estimation respectively, cooperating with gain functions computed by spectral subtraction (SS) and log-spectral amplitude (LSA) algorithms. And we will compare the purification results under musical noise and noise from freeway.