G. Rogozinsky

Among the variety of existing perceptual audio coding algorithms, MPEG-1 Layer 3 also known as MP3 is considered to be the most popular because of its simplicity and efficiency. However, some problems exist here, mostly related to inaccuracy of psychoacoustic model used in MP3. The coder utilizes bank of 32 polyphase band-pass filters with equal bandwidth to approximate critical bands of human ear. This approximation doesn't correspond well to results of multiple psychoacoustic experiments, which show that the width of critical bands rises with central frequency increasing. One of the solutions for this problem is to exploit the wavelet-based psychoacoustic model. The wavelet packet transform produces an adaptive orthogonal domain and divides input signal according to chosen tree structure. Each time the tree splits the signal passes two quadrature mirror filters with impulse responses correspond to scaling and wavelet functions. Thus it divides signal in frequency domain into several bands with different width. Such model approximates critical bands of ear with better precision comparing to MP3 psychoacoustic model. Unfortunaly, the frequency selectivity of wavelets is not enough for high-quality audio processing when precise frequency separation is needed. This fact limits application of wavelets in audio coding. In reviewed paper wavelet optimization method is discussed, based on modified Remez exchange algorithm. The following method allows obtaining the new wavelet bases with improved frequency properties mainly because of decreasing the number of vanishing moments. The author observes attenuation in stopband of wavelet filter under different optimization parameters such as filter length, number of vanishing moments, transition band width. Additionally, when optimizing wavelets, the chosen tree structure is considered to be optimal for approximation of critical bands by wavelet packets.