用於音訊壓縮之小波及弦波分析

Abstract

小波和弦波表示法對高效率音訊編解碼而言是項深具潛力的工具。然而這兩種技術在音訊編解碼上的研究才剛剛起步，有許多地方仍值得進一步探索。這篇論文的目的主要在結合弦波分析和小波轉換兩種技術去建構一個高效率的音訊壓縮系統。首先，我們採用線性預測的技巧以找出存在於音訊資料內的弦波成分。單一音訊資料框有時需要被切割成兩個獨自擁有一套表示法的資料框，以增加弦波表示法的準確度。當音訊訊號內的弦波成分被移除後，剩餘的部分將會進行小波封包轉換。藉由跨頻帶預測，我們試著減少子頻帶間重複的資訊，且在最後輸出已被感觀式量化編碼的資料。在上述的演算法中，存在著許多細節技術。例如，在弦波分析中，我們需要一個適當的對照程序來消除不正確的頻率估計。另外，在決定資料框分割處時，我們必須對不同規模的小波轉換係數作仔細的判斷評估。為了驗證論文內提到技術的可行性，我們將進行各別項目的模擬並提出報告。其中有些技術方法顯示出其對編解碼確實有幫助，但有些方法則需要更進一步的研究來釐清其是否適合使用於音訊壓縮。

Wavelets and sinusoidal representations are two promising tools for high efficient audio coding. However, the study of these two techniques on audio coding just started. Many unknown areas remain to be explored. This thesis is an attempt of combining both sinusoidal analysis and wavelet filter together to form an efficient audio compression system. First, we adopt a linear-prediction technique to estimate the tonal components in an audio sequence. To increase the accuracy of sinusoidal representation, an audio frame may be split into two segments, each containing a set of sinusoids. After the sinusoids are removed from the audio sequence, the residues are passing through a set of wavelet packets. Then, the redundancy among subbands are reduced by inter-subband prediction. The final outputs are quantized perceptually. There are a number of detailed steps in each technique of the preceding algorithm. For example, in sinusoidal analysis (extracting sinusoidals), a proper screening process is needed to eliminate incorrect candidates. Another novel technique, audio frame segmentation, requires a careful check through different layers of wavelet packet outputs. To examine the effectiveness of the techniques employed in this thesis, numerous simulations have been conducted and reported. Several techniques are shown to increase the coding efficiency, while the others may need further study to clarify their usefulness.