MPEG1 Layer I, II 架構下之立體音訊壓縮探討

Abstract

立體音訊壓縮的目的是為了減少位元率，並且在解壓縮後仍然保證音訊 的品質。MPEG1 是目前在各種商業應用上最被廣泛使用的數位音訊壓縮標 準。MPEG1 第二層在對雙聲道音訊個別獨立壓縮的情況下，可以達到在 2x128 kbits/s以上的位元率有如同原CD品質的音效。進一步使用雙聲道 合併編碼(joint stereo coding)的方法，例如MPEG1中的雙聲合一編碼 法(intensity stereo coding)，在2x128 kbits/s以下的位元率可以有較 佳的音訊品質。在本論文中，將分析立體音訊中雙聲道資料的重複性( redundancy)。在MPEG1 第一、二層的架構下，我們利用Karhunen-Loeve 轉換及雙聲道間的線性預測這兩個方法來分析及利用立體音訊的資料重複 性。我們利用Karhunen-Loeve轉換提出兩種改良的MPEG1雙聲合一編碼法 ，改善低於2x128 kbits/s的位元率下的解壓縮立體音訊品質。主觀及客 觀測試的結果說明了我們提出的改良MPEG1 雙聲合一編碼法較原方法有較 佳的立體音訊品質。在雙聲道間的線性預測方法上，我們考慮各種不同的 參數，如預測的階層(prediction order)、延遲的參數、隨時間的改變性 、額外的訊息(side information)等，來分析壓縮增益(coding gain)。 實驗的結果建議雙聲道間的線性預測方法在實際的應用上應使用在訊號低 頻的部分，並且每經過多個編解碼框才傳遞一次線性預測係數以降低額外 的訊息傳輸。 The purpose of stereo audio signal coding is to reduce the required bit rate, while maintaining the signal quality after decoding. The ISO MPEG1 is the most widely used audio compression standard in many commercial applications. Among the vast commercial products, MPEG1 layers I and II coding processes are most widely adopted. MPEG1 layer II can achieve a transparent audio quality above 2x128 kbits/s by independent coding of the left and the right channels. With the use of joint stereo coding technique, such as intensity stereo coding in MPEG1, the decoded audio quality can be improved for the bit rate lower than 2x128 kbits/s. In this thesis, we analyze the data redundancy of stereo audio signals. The Karhunen-Loeve (KL) transform and inter-channel prediction methods are applied to exploit and analyze the data redundancy in the framework of MPEG1 layers I and II. On the KL transform, we propose two modified intensity stereo coding algorithms for MPEG1 layers I and II by KL transform to further improve the decoded stereo audio quality at bit rate below 2x128 kbits/s. Subjective and objective measurements show that the two algorithms have better stereo audio quality than the original MPEG1 method. On the inter-channel prediction, we consider the coding gains along with various parameters such as prediction order, prediction delay, time varying property, the required side information, etc.. The experiment results suggest the applying of inter- channel prediction in the low frequency bands, and transmission of the prediction coefficients once for longer frames to avoid the side information overhead.