Dear Stefan,
In the paper "Improved Forward-Adaptive Prediction for MPEG-4 Audio Lossless
Coding", a non-linear compander is applied to the parcor coefficients prior to
quantization. This compander is designed in order to minimize quantization
error, especially for magnitudes close to unity.
If you determine the typical distribution of magnitudes of the LPC
coefficients, you could design a good non-linear compander in order to minimize
the error due to LPC coefficients quantization. By doing this, the input signal
would be better estimated and the output residue (which occupies more than 90 %
of the output bitrate) would be smaller.
Regards,
Fernando A. Marengo Rodriguez
Acoustics and Electroacoustics Laboratory
Faculty of Sciences, Engineering and Surveying
National University of Rosario
Rosario, Argentina
http://www.fceia.unr.edu.ar/acustica/codecdigital/integrantes.html#Fernando
Message: 1
Date: Tue, 19 Jul 2011 14:48:27 +0200
From: Stefan Westerfeld <[email protected]>
Subject: [Flac-dev] Using line spectral pairs for LPC quantization
To: [email protected]
Message-ID: <[email protected]>
Content-Type: text/plain; charset=us-ascii
Hi!
I've recently read the flac sources, and it appears to me that you're quatizing
the LPC coefficients directly. However, this might be not as efficient, because
LPC coefficients are very sensitive to quantization error. I don't know how
much space the LPC coefficients occupy in the resulting FLAC file - if its not
much, it probably doesn't matter.
But if its a significant amount of data, it might be better to convert the LPC
coefficients to LSF/LSP coefficients, and store those quantized. This should
provide the same quality at a lower bit count, if I'm right. Here is a link
that describes LSF/LPC coefficients:
http://en.wikipedia.org/wiki/Line_spectral_pairs
Cu... Stefan
--
Stefan Westerfeld, Hamburg/Germany, http://space.twc.de/~stefan
_______________________________________________
Flac-dev mailing list
[email protected]
http://lists.xiph.org/mailman/listinfo/flac-dev
