Optimum Quantization of Flattened MDCT Coefficients*

摘要

The performance of perceptual audio coders depends on the efficiency of the quantization operation in masking the quantization noise under the audio signal. This objective is better addressed by coding separately different signal components such as sinusoids, transients and stationary noise. In this paper we use an audio coder that normalizes the MDCT spectrum by a smooth spectral envelope and by periodicities due to sinusoids. The resulting flattened MDCT coefficients are shown to exhibit a probability density function with small uncertainty allowing the design of an optimum non-uniform scalar quantizer. Its distortion-rate function is derived, is compared to that of known quantizers, and is compared to that obtained under real audio coding conditions.