使用DPCM-LMS串接預估器之無失真音訊編碼

LOSSLESS AUDIO CODING USING CASCADED DPCM-LMS PREDICTION

李鑑哲(Chien-Che Lee)

指導教授 ： 李清坤

大同大學/電機資訊學院/通訊工程研究所/碩士(2008年)

摘要

無失真音訊編碼(lossless audio coding)的壓縮效能雖然顯著低於失真音訊編碼(lossy audio coding)，但是由於它可以將壓縮後的音訊作完全相同的還原，也就是解壓縮成原始的音訊，因此無失真音訊編碼在某些應用上有其必要，例如高品質的音樂保存，專業音樂創作者的需求，及保存具有歷史意義的語音或是聲音。無失真音訊壓縮所使用的預估器(predictor)可以分成區塊(block based)和逐點(sample-by-sample based)兩種不同的方式。逐點式預估器由於在通訊應用上有低延遲(low delay)的特性，故有其實用的價值。而逐點式預估器最常使用的演算法是適應性最小均方演算法(adaptive least mean square, adaptive LMS)。為了讓無失真音訊編碼得到較好的效能，適應性預估器通常以串接(cascade)且高階(high-order)的方式存在，但是此方式的運算量比其它類型的預估器還要大很多。此篇論文的目的在於設計低運算量的串接預估器，藉由測試各種預估器的效能，尋求比單一相當階數預估器效能更佳且運算較少的串接型預估器。
為了尋找較適當的固定型預估器和低階的最小均方預估器做串接，我們以遞迴減法的方式尋找無需乘法的固定型預估器和低階的最小均方預估器作串接，取代對其階數的增加，以達到減少預算量或是效能的增加。會以低階的最小均方預估器為對象的原因是當其為較高階時，其效能不一定會增加。經過初步的測試結果，將固定型預估器置於最小均方預估器之前，會有較好的效能，這也是本研究所採用的方式。而這邊所使用的固定型預估器除了無需乘法外，且細分為兩種方式和最小均方預估器作串接，一種為單一的固定型預估器直接和最小均方預估器串接，另外一種串接的方式為以區塊的方式從多個不同階數的固定型預估器中選出產生最小絕對值誤差和為其所要的預估器。
最後的測試結果顯示出單一的差分脈衝編碼(Difference Pulse Coding, DPCM)比其它大部分無需乘法器的固定型預估器，包括以區塊的方式從多個不同階數的固定型預估器以誤差條件選出的預估器和其它單一直接和最小均方預估器做串接的固定型預估器，還要適合和低階數的最小均平方預估器互相串接，本研究的測試結果也發現在四階以下，差分脈衝編碼和最小均方預估器串接成的預估器或是單一的最小均方預估器對其增加階數確定都會有效能的提升，而差分脈衝編碼和最小均方預估器串接成的預估器，當其最小均方預估器為二階的時候比單一的最小均方預估器為三階的時候會有更好的效能。

關鍵字

最小均方演算法 ； 預估器 ； 無失真音訊編碼

並列摘要

Although the compression performance of lossless audio coding is significantly worst than lossy audio coding, lossless audio coding can recover the original audio signal from the compressed audio signal, that is compressed audio signal can be decompressed to the original audio signal, for this reason lossless audio coding is necessary for applications, such as storage of high quality music, professional music maker, reserve of historical speech or sounds. There are two kinds of predictors for lossless audio coding, one of them is block based, and the other is sample-by-sample based. The most common sample-by-sample based prediction is adaptive prediction, due to the low delay of adaptive prediction, adaptive prediction is worth to be used in communication applications and adaptive least mean square (we simply use least mean square, LMS, in the entirely following texts) is the most common used sample-by-sample prediction. In order to get better performance, adaptive prediction is used in cascaded form and each stage is at high-order mode, leading to much higher computational requirement than other kinds of prediction. The purpose of this research is to design a low-computation cascaded predictor. By testing various kinds of predictors, we try to find a suitable cascaded predictor having higher performance and lower computation than a single equivalent- order LMS predictor.
In order to find the suitable cascaded fixed-LMS or LMS-fixed prediction, we use the recursive subtraction method to get the cascaded multiplicationless-fixed prediction and low order LMS prediction instead of increasing the order of LMS to achieve decreasing the computational quantity or improving performance. The reason of using low-order LMS is if the order of LMS is high enough, the higher-order LMS doesn’t mean improving performance. Since the pre-testing of fixed-LMS/LMS-fixed prediction, we found fixed-LMS prediction would get better performance, we decided to use the form of fixed-LMS in the entire experiment. The fixed prediction here are two types, first is the single fixed predictor, second is to choose a block based fixed predictor which generates minimum summation of absolute error from a set of fixed predictors.
The final testing results reveal that single DPCM predictor is more suitable than other multiplicationless-fixed predictors which are including block based fixed prediction with error criterion and single fixed predictor to cascade with single low-order LMS predictor. We also found that DPCM-LMS/LMS prediction would get performance improvement when the order is increasing between order-1 to oreder-4 and DPCM-LMS prediction at order-2 has slightly better performance than single LMS prediction at order-3.

並列關鍵字

lossless audio coding ； adaptive least mean square ； predictor

參考文獻

[2] H. Huang, P. Franti, D. Huang, S. Rahardja, "Cascaded RLS–LMS Prediction in MPEG-4 Lossless Audio Coding," IEEE Transactions on Audio, Speech, and Language Processing, vol. 16, no. 3, pp. 55-562, Mar. 2008.

[9] P. Prandoni and M. Vetterli, "An FIR Cascade Structure for Adaptive Linear Prediction," IEEE Transactions on Signal Processing, vol. 46, no. 9, pp. 2566 – 2571, September, 1998.

[10] Lossless Compression of Audio, [Online]. Available: http://www.firstpr.com.au/

[12] M. Hans, R.W. Schafer, "Lossless Compression of Digital Audio," HP Laboratories Palo Alto, HPL-1999-144, Nov. 1999.

[13] D.Y. Huang, X. Su and A. Nallanathan, "Characterization of a Cascaded LMS Predictor," IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2005, pp. 173-176.