Quantization in perceptual audio coders with correction for synthesis filter bank noise spreading

Perceptual transform and subband audio coding systems using analysis/synthesis filter banks introduce noise by quantizing subband signals or transform coefficients. Historically, systems have been designed under the assumption that the quantization noise power spectra at the synthesis filter bank input and output are equal. In general, the power spectra will have different shapes since the synthesis convolution leads to quantization noise spreading. In this paper, a theoretical framework for deriving an optimum noise allocation that accounts for spreading by the synthesis filter bank is described. In concept, the problem of finding an optimal noise allocation can be expressed as a constrained optimization problem. A simplified algorithm derived from the theoretical framework is discussed that was found to improve subjective audio quality while using only modest computational resources.