Paper Info
Title
Optimization-quantization for least squares estimates and its application for lossless audio compression
Abstract
In this paper we study the problem of optimally quantizing the least square estimates and we introduce a method where the quantized es- timate vector is obtained by a sequence of interleaved optimization- quantization scalar stages. We show how the general approach can be reduced to a simple and ef� cient algorithm when connecting it to the LDLT solver for general LS problems. The application to quan- tization of the linear prediction coef� cients for audio lossless coding reveals the high performance of the approach, leading to topmost performance in the class of frame based audio coders, surpassing signi� cantly the performance of the current MPEG4-ALS standard. The quantization of LS estimates and tracking the loss of perfor- mance due to quantization is a generic problem which can be en- countered in all areas of engineering and science. We consider the application of least squares linear prediction for asymmetrical loss- less audio compression, where the prediction coef� cients are trans- mitted as side information, making decoding very fast. The study of quantization of linear prediction coef� cients (LPC) has a long history and we can distinguish two distinct areas of applications: the� rst is lossy compression (including the important application to speech coding) and second is frame-wise (or forward) lossless com- pression. Here we use the later technique and we note that there are many equivalent representations of LP coef� cients, such as re�Ä ec- tion coef� cients, log-area ratios of re�Ä ection coef� cients, and arcsine of re�Ä ection coef� cients. Scalar quantization can be applied to any of these representations, obtaining different results in the� nal com- pression application. We note that the conclusions of most previous studies were favoring always the alternative representations and con- sequently the quantization of the direct representation of LPC was traditionally considered only a bad choice. However, we are going to show that working with the quantized direct form of LPC we get remarkably good tradeoff predictor complexity-prediction accuracy, and having this technique implemented in an audio codec provides the best performance available in terms of compression ratios and encoding/decoding times, surpassing all existing methods and stan- dards in the� eld.
Year
DOI
Venue
2008
10.1109/ICASSP.2008.4517579
Las Vegas, NV
Keywords
Field
DocType
audio coding,least mean squares methods,quantisation (signal),audio lossless coding,interleaved optimization-quantization scalar stages,least squares estimates,lossless audio compression,MPEG4-ALS standard,least squares,linear prediction,lossless audio compression,optimal quantization
Least squares,Computer science,Theoretical computer science,Audio Lossless Coding,Artificial intelligence,Pattern recognition,Algorithm,Linear prediction,Quantization (physics),Solver,Data compression,Quantization (signal processing),Lossless compression
Conference
ISSN
ISBN
Citations 
1520-6149 E-ISBN : 978-1-4244-1484-0
978-1-4244-1484-0
2
PageRank 
References 
Authors
0.46
1
2
Name
Order
Citations
PageRank
Florin Ghido1356.97
Ioan Tabus227638.23