使用卡曼濾波器追蹤參考訊號之適應性語音純化波束形成器

Abstract

本論文提出一套利用麥克風陣列來降低噪音及迴響效應的演算法。在實際環境中，目標音訊不只常受到穩態雜訊及非穩態雜訊的干擾，更常因為迴響效應而使語音品質遭到破壞。因此，本論文期望設計一個能濾除雜訊並減少迴響影響的適應性波束形成器。此演算法在波束形成器演算法中，引入參考訊號的觀念並輔以Kalman濾波器來進行演算。此外，經過些微的修改，本演算法也可以利用於偵測語音活動。藉由適當的語音活動偵測，可以幫助分辨目標與噪音在本質上的不同，並且加速Kalman濾波器的收斂。利用實際在車上錄得的音檔進行的實驗結果也在此篇論文中呈現。本論文並利用客觀的參數評估所提出的波束形成器與語音活動偵測的效能，並與其他已知的方法進行比較分析。

In this thesis, an algorithm that considers noise reduction and de-reverberation simultaneously using microphone array is proposed. In many practical environments, the desired speech signal is usually contaminated by stationary or non-stationary noises and distorted by reverberation. When considering noise reduction only, the desired speech signal could be distorted further due to the effect of desire signal cancellation etc. The objective of this thesis is to design an adaptive beamformer to incorporate de-reverberation into the noise reduction framework. The proposed method tracks a pre-recorded reference signal to compensate the reverberation effect. Consequently, the algorithm results in a trade-off between the two objectives. Further, a voice activity detection (VAD) algorithm is proposed by slightly modifying the proposed algorithm. An adequate VAD can help to identify the nature of signal and noise and accelerate the convergence rate of Kalman filter. The experiments on real car sound samples are processed. The performance of beamformer and voice activity detection are both evaluated and compared with existing algorithms.