Perceptual Coding of Audio Using Signal - AdaptiveFilterbanks

This thesis studies the application of signal-adaptive lter banks in perceptual coding of audio with an emphasis on Wavelet Filter Banks (WFB). It provides an overview of perceptual coding of audio, the motivating psychoacoustic principles, transforms and wavelet theory. Additionally, di erent existing wavelets-based audio coding schemes are presented. The aim of most of the schemes is to overcome the poor frequency localization of the wavelet transform. Finally, a simple MLT-based coder and a signal-adaptive wavelet-based coder are implemented and tested. The proposed wavelet coder employs a hybrid lter bank where the rst stage uses a Discrete Wavelet Transform (DWT) and the second stage uses one or more Discrete Cosine Transforms (DCT) of variable length and the coding gain is used as switching criterion. Two adaptive bit allocations are available. One method uses a theoretically derived relation that assumes a uniform distribution of the quantization error and the other one uses an operational rate distortion measure. The samples are quantized with a linear scalar quantizer and the rate is estimated via entropy. The coder is tested for di erent wavelet lter sets as well as for di erent DCT lengths. It turns out that the adaptive DCT stage yields a clear improvement especially for harmonic signals like piano music that have shown to be problematic for wavelet coders. 2 CONTENTS