Audio Signal Processing and Coding

6.2 ANALYSIS-SYNTHESIS FRAMEWORK FOR M-BAND FILTER BANKS

Filter banks are perhaps most conveniently described in terms of an analysis-synthesis framework (Figure 6.1), in which the input signal, s(n), is processed at the encoder by a parallel bank of (L − 1)-th order FIR bandpass filters, H_k(z). The bandpass analysis outputs,

are decimated by a factor of M, yielding the subband sequences

which comprise a critically sampled or maximally decimated signal representation, i.e., the number of subband samples is equal to the number of input samples. Because it is impossible to achieve perfect “brickwall” magnitude responses with finite-order bandpass filters, there is unavoidable aliasing between the decimated subband sequences. Quantization and coding are performed on the subband sequences y_k(n). In the perceptual audio codec, the quantization noise is usually shaped according to a perceptual model. The quantized subband samples, , are eventually received by the decoder, where they are upsampled by M to form the intermediate sequences

In order to eliminate the imaging distortions introduced ...