Investigation And Improvement Of MPEG-2 AAC Codec

Low bit-rate perceptual audio coding exploits perceptual irrelevancies and statistical redundancies to compress audio signals, aiming at achieving transparent audio quality with a minimum number of bits. MPEG-2 AAC uses the two coding strategies mentioned above to significantly reduce the bit rate used to represent audio signals. It has been adopted as the core standard in the audio coding part of MPEG-4. Moreover, MPEG-2 AAC is regarded as the current state-of-the-art technology of audio compression, and it is designated as the MP3's successor. Firstly, perceptual coding principles are introduced in this paper; The psycho-acoustic model is described in detail. Then, an in-depth study on the primary modules of AAC is presented. Finally, improvements on several algorithm parts involved in AAC standard are described. In the standard AAC, scalefactors are used to shape quantization noise. Via a nest of two iteration loops, the best tradeoff between the bit rate and audio quality is found. Scalefactors are encoded with Huffman coding. Huffman coding involves codebook design, search and storage. It may incur high computational complexity and storage requirement, especially when codebook size or number of symbols is large. As it is known, arithmetic coding outperforms Huffman coding in some aspects. Therefore, we make attempts to encode scalefactors with adaptive arithmetic coding instead of Huffman coding. The objective of our effort is to reduce bit rate while maintaining audio quality. Performance of adaptive arithmetic coding depends on accuracy of the involved statistical model. In general, adaptively statistical model becomes accurate when the coded symbols are enough. In order to improve the statistical accuracy, we proposed an adaptive model based on context and features of audio coding. Using our proposed model, the scalefactors of current frame are represented with the statistical distribution of the N preceding frames. Finally, we implement AAC encoder and decoder and their improved versions on a X-86 platform. What we have done may offer a solid foundation for further research and applications.