Optimal Quantized Linear Prediction Coefficients for Lossless Audio Compression - Scalar Quantization Revisited
Uniform scalar quantization of linear prediction coefficients with B bits is traditionally done by multiplying each coefficient with Q = 2B and rounding it to the nearest integer. We propose an improved, optimal quantization method by replacing the rounding with a more elaborated procedure. It uses on average 2 bits less per quantized prediction coefficient for a similar sum of squared errors and allows an accurate estimate of the mean squared error misadjustment as a function of Q for a given subframe and predictor order M. We introduce several efficient time-constrained probabilistic search methods for obtaining near optimal solutions. There are no required changes at the decoder and the method is applicable on a wider area of cases (mono, stereo, and multichannel prediction) than quantization of reflection coefficients. Moreover, the method enables near optimal compression for 24 bit audio using only 32 bit arithmetic operations.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is temporarily free for AES members.