MPEG-4 HE-AAC音訊壓縮標準之研究及其在Trimedia DSP上之實作探討

Abstract

最新的音訊編碼標準, MPEG-4 HE-AAC, 可以利用Spectral Band Replication (SBR)的技術達到很高的壓縮效率。和傳統的AAC 相比,使用SBR之後可以節省大約30% 的位元率。 然而, 由SBR 造成的如此的效率提升造成了計算複雜度的大量增加。 在越來越多的多媒體應用著重於攜帶式裝製或低價多媒體處理器上, 對低功率設計的需求成為一個重要課題以致於HE-AAC 的高計算複雜度較無法被這些應用所接受。 在這份論文裡, 我們在Trimedia DSP 晶片上實作了一個HE-AAC 解碼器並且使用數種演算法來改進程式的效率。此外, 我們應用一種評估程式功率的工具來證實我們所採用的演算法能夠節省解碼器的功率消耗。根據運算時間的分析結果顯示, SBR 裡的Quadrature Mirror Filters(QMF) 這個部分花費HE-AAC解碼器大約70%-80% 的執行時間。為了減低QMF 的計算複雜度, 我們可以採用實係數的filterbank (LP-SBR) 。再透過一些filterbank 分解方法我們可以進一步加速QMF 的執行過程。另外, 用16 位元定點數來表示QMF 的取樣點不僅可以藉由更快速的定點數指令來增加執行效率，還可以藉此充分運用Trimedia 訊號處理器固有的平行性特徵。以Trimedia 超長指令字(VLIW) 的結構和單一指令多重資料(SIMD) 的指令集, 每二個QMF 取樣點可以在單一的指令內運算完成來加速QMF。儘管由LP-SBR 和定點數逼近造成的音質的損失人耳幾乎無法察覺, 我們仍然使用了一些增進音質方法來補償這些降低的音質。總結這些提出的演算法, HE-AAC 解碼器可以在維持高音質的情況下達到很高的執行效率。

The state-of-the-art audio coding standard, MPEG-4 HE-AAC, achieves high coding efficiency by the employment of Spectral Band Replication (SBR) technique. Comparing with conventional AAC, about 30% of the bit-rate is saved with the use of SBR. However, the improved performance contributed by SBR results in a huge increase of computational complexity. With more and more multimedia application focus on the portable devices or low-cost media processors, the demand for low-power design becomes an important issue so that the high computational complexity of HE-AAC is undesirable. In this thesis, we implement an HE-AAC decoder over the Trimedia DSP chip and employ several algorithms to improve the performance. Besides, a power evaluation tool is also used to testify that the algorithms adopted do decrease the power dissipation of the decoder. According to the profiling analysis, Quadrature Mirror Filters (QMF) used in SBR takes about 70%-80% execution time on HE-AAC decoder. To reduce the complexity of QMF, a real-valued filterbank (LP-SBR) is first applied. We further accelerate the QMF process via filterbank decomposition methods. Besides, 16-bit fixed-point representation of QMF samples are applied not only to increase the performance by faster operations but also to make it possible utilizing the inherent parallelism characteristics of Trimedia DSPs. With Trimedia Very Long Instruction Word (VLIW) architecture and Single Instruction Multiple Data (SIMD) instruction set, every two QMF samples can be operate in a single instruction to accelerate the QMF process. Though the quality loss caused by LP-SBR and fixed-point approximation is nearly imperceptible, a quality enhancement method is applied to compensate the quality degradation. With the proposed schemes, the HE-AAC decoder achieves high coding performance with still high sound quality.