雙麥克風回音壓抑之研究

Abstract

通常降低迴音的方式有兩種，包括迴音消除與迴音壓抑。迴音消除的估計對象是迴音，直接地消除接收端的迴音，但其效益易受近端說話所影響；迴音壓抑的估計對象是迴音功率，透過頻譜資源重新分配的方式，凸顯訊雜比較好的頻段，壓抑訊雜比較差的頻段，間接地降低迴音對接收端的影響，雖然抗近端說話的能力比迴音消除強，但付出了近端失真的代價。隨著科技進步，人們對語音享受的需求增加，立體聲逐漸普及，立體聲迴音處理也逐漸受到重視。在迴音不均等下，修正了傳統依序式估計，也提出分開式、加總式、單邊式估計的方式。近端會對使迴音功率的估計誤差與輸出的訊雜比增加。在雙麥克風模型下，我們針對近端的兩個效應提出三種改善模型，如降低近端干擾、提高輸入訊雜比與混合式模型，當兩迴音處理系統組合時，迴音壓抑受近端說話的影響增大，我們可以把迴音消除所估計的迴音當作新的估計器訊號，使迴音壓抑能更準確地估計殘餘迴音。

Echo reduction model includes acoustic echo cancellation (AEC) and acoustic echo suppression (AES). AEC cancels echo directly by estimating echo, but it is easily affected by double talk. AES suppresses echo indirectly by estimating echo power. It increases signal-to-noise-ratio by different spectral gains at the cost of distorted signal. In case of stereo echo imbalance, traditional sequential AES model is modified. We also propose separate model, sum model and single model. Near end signal can increase echo power estimation error and increase output signal-to-noise-ratio. To overcome its difficulties, we add a secondary microphone to establish two-microphone AES model. By substraction of properly delayed microphone signals, we also propose three different methods to reduce interference of near end signal. Finally, AEC and AES are combined to enhance echo reduction performance.