噪音消除演算法的設計與實現

Abstract

本論文將提出一項用來改善在噪音環境之下通話音質的單聲道噪音減除演算法，此演算法比起過往的技術在造成音質扭曲上的問題有更佳的表現。此系統主要包含了兩大部分：線性預估編碼系統以及Wiener濾波器系統。相較於過去常用的Wiener頻域減除方式所造成擾人的音樂噪音問題，所提出的演算法利用動態估計噪音大小的方式來實現新的噪音減除函數。除此之外，以線性預估編碼為技術的前處理器被用來萃取出聲音訊號中含有頻率組成的人聲部份。這一道處理是保留備處裡過後聲音訊號音質的重要步驟。所提出的演算法被實現在定點運算的ＤＳＰ板上並且以主觀的聆聽測試來證明其效果。包含白噪音以及有染色的冷氣和吹風機噪音都被用來作為測試訊號。其色次結果則用變異數分析方法來做為分析的工具。比較起傳統慣用的噪音減除演算法，新提出的方法在改善含有噪音的語音訊號上面的效果有明顯的進步並且提供更棒的音質。

An efficient algorithm is presented in this thesis for single-channel noise reduction (NR) without impairing timbral quality. The system mainly consists of two function blocks: the linear prediction module and the Wiener filtering module. Unlike conventional Wiener spectral subtraction methods that suffer from disturbing musical noise, the proposed method utilizes an improved approach for estimating dynamically the noise power. In addition, a preprocessor is used to extract the correlated structure of human speech by using linear prediction coding (LPC). This preprocessing step is vital to preserve timbral quality and spectral envelope of the signal. The proposed algorithms were implemented on a fixed-point digital signal processor (DSP) and validated using subjective listening tests. White noise and colored noise from an air conditioner and a hairdryer are used as test noises. The test results were processed using Analysis of Variance (ANOVA). As compared to conventional NR algorithms, the proposed algorithm proved remarkably effective in enhancing noise corrupted speech signals, without compromising the timbral quality.