二階式聲學迴音消除系統之研究

Abstract

本論文首先介紹聲學迴音消除技術，接著提出二階式迴音消除系統的架構，其效能在於可以使用較少的運算量，而達到傳統單階消除系統的效果。系統前級使用長度較短的傳統NLMS濾波器過濾大部份迴音，而後級的迴音塑型濾波器分別使用時域的威能濾波器與頻域的頻譜消減濾波器進行迴音殘響消除。時域系統運算簡單，藉由兩階段處理可大幅減低計算量，但在雙方通話狀態無法兼顧迴音消除與保持近端語音完整。在頻域系統中，藉由估測迴音殘響的頻譜能量再加上聲響心理學上之遮蔽效應，在頻域上只需消除被人耳察覺的迴音，使近端語音能保持最小失真，真正達到提高通訊品質的目的。

In this thesis, we present the basic formulations of a two-stage acoustic echo canceller. It offers sufficient echo reduction with less computational complexity. In the proposed system, we use a NLMS FIR filter to cancel the early reflection echo, followed by an echo shaping filter to reduce the echo residue. The echo shaping filter can be realized using a time-domain Wiener filter or a frequency-domain generalized spectral subtraction filter. While the time-domain filter is easy to compute, it has some limitations as far as the double talk is concerned. When applying the frequency-domain filter, power spectral densities of echo residue are estimated and then subtracted from the compensation signal. We also explore the benefits of incorporating masking effect of phychoacoustics in the frequency-domain system. Simulations results indicate that the combined use of echo shaping filter and the masking threshold provides an ideal compromise between the echo cancellation and near-end speech quality.