Audio Coding Using Variable-Depth Multistage Quantization

摘要

In this paper, an algorithm for high quality coding of 48 kHz sampled audio signals is presented. The signal is first converted to a frequency-domain representation using the Modulated Lapped Transform. Perceptual irrelevancies in the signal are removed by employing frequency-domain human auditory masking properties to derive a masked quantization noise threshold. The noise threshold is then smoothed using simple morphological operations to account for limitations in the frequency resolution of the transform. This smoothed threshold is then used as the input to a multistage uniform scalar quantization scheme that shapes quantization noise to fit below the masking threshold. Higher order entropy coding is applied to the output of the quantizer, simultaneously exploiting dependencies in the time, frequency, and stage dimensions. A variable number of quantization stages is used for each transform coefficient. However, the nature of the formulation is such that no side information is required to send the number of stages. Through an informal experiment, the audio quality of the new coder was found to be better than that of the MPEG layer 1 coder and roughly equivalent to that of the NIPEG layer 2 coder.