基於一個線性R-D模型的快速音訊位元分配技術

Abstract

MPEG-1 Layer III 音訊編碼方式是目前網際網路上最被廣泛使用的數位音訊格式，然而由於其編碼器的複雜性較其他層次(Layer)高，且其位元分配模組在整體運算時間中,佔有很大的比例，因此本篇論文利用一個線性模型提出一個快速的位元分配技術，在Rate Control Loop中，大幅地減少迴圈運算次數；而在Distortion Control Loop中，利用估測量化誤差的技術，以改良的NMR(Noise-to-Masking Ratio)平行法，運用單次迴圈計算出所要的scalefactors。我們所提出的位元分配技術和目前網際網路上知名的MP3編碼軟體【LAME】比較，約有五倍左右的加速。在客觀的壓縮音訊品質測量中，我們採用了三種方式來評比，分別是整體的失真頻帶(scalefactor bands)數、平均的整體NMR和客觀差異分數(Objective difference grade) [6]。我們的實驗結果顯示出我們所提出的理論能有效顯著地增加編碼速度且仍可維持一定的音訊品質而無明顯的失真發生。

The MPEG-1 Layer III audio coding is most widespread digital audio format on the Internet. Its encoder is more complicated than the other layers in the MPEG-1 audio specification, where the bit allocation constitutes significant part of the total computational load. This thesis describes a fast bit allocation technique using a novel linear model to compute the bitrate and the global gain of the inner loop (rate control loop) for the classic MP3 bit allocation problem. Estimated quantization noise is approximated with a binomial expansion for the adjustment of the scalefactors. This procedure reduces the outer loop iterations (distortion control loop) to one loop. Our bit allocation technique is about five times faster compared to the well-known LAME MP3 encoder. Based on the objective quality measure, total number of distorted bands, average total noise-to-masking ratio (NMR), and objective difference grade [6], our experimental results show that there is significant improvement in speedup of the encoding process without noticeable audio quality degradation.