基於頻譜形狀敏感度的感知聲學迴音消除演算法

Abstract

傳統用於聲學迴音消除系統的效能指標如ERLE、MSE只有從數學觀點上做量化的評比，而沒有考慮殘餘迴音的可聽性。 人耳聽覺除了對於各種頻率有不同的敏感性以外，大腦也透過時-頻域的調變特性來增進對語句的了解。而這樣時-頻域調變濾波處理的概念也已經廣泛的用於許多語音處理的應用中如雜訊消除、語音動態偵測和語音可讀性的量測。

在本論文中，我們利用聽覺閾值和頻域調變的感知閾值設計了二階的預強調濾波器用於NLMS演算法，並使AIR (Aachen impulse response)資料庫去模擬真實的室內迴音環境並利用TIMIT語料庫進行測試。最後經由聽覺實驗證明這樣的設計能增進聽覺較為敏感頻帶的收斂速度，並使殘餘迴音在收斂過程中變得較為小聲。

Conventional acoustic echo cancellers are developed and evaluated using mathematical perspective performance measures. They are seldom evaluated from perception viewpoints by assessing the perceived residual echo. In addition to the fact that human hearing possesses different sensitivity to different frequencies, our brain analyzes sounds in terms of their spectro-temporal modulation contents. This spectro-temporal modulation filtering concept has been successfully used in many speech-related applications, such as noise reduction, voice activity detection and speech intelligibility assessment.

In this thesis, we design a two-stage pre-emphasis filter based on hearing thresholds and detection thresholds of the spectral modulation transfer function for NLMS acoustic echo cancellers. The Aachen impulse response (AIR) database and TIMIT corpus are used to simulate real acoustic echo environments. Listening tests show that our design, which increases the convergent rates in those hearing-sensitive frequency bands and spectral modulation bands, effectively reduces the perceived residual echo produced by echo cancellers during their convergence.