[vorbis] Cascading?

[Gregory Maxwell]

On Tue, Feb 27, 2001 at 06:37:52PM +0100, Aleksandar Dovnikovic wrote:
> During the interesting interview that binaryfreedom has made with Monty
> and Jack, Monty mentions cascading, a feature that will be added, quote:
> "Cascading is the ability to make multiple passes through the frequency
> spectrum, iteratively filling in more detail, like a progressive jpeg".
> What are the advantages of something like this - does this generally
> improves quality or is it used primarily to put more important audio data
> at the beggining of the block so that bitrate peeling might be effectively
> applied?

1. Yes, it's better for bitrate peeling.
2. It lets you use higher dimensionality without having corseness in your
error, or in the case of non-lattic books, without sending a gigantic
codebook to accuratly represent your data.

Lets pretend that there is a high simmlarity of your Floors on a sample
set (perhaps your sample set is a book on tape, a clean recording of a
single male voice). Only a small part of the 32 dimension floor lsp space
will be used by this set.

If you encode by small vectors of LSP (break it up into 2 or 4 lets say) you
will not express this redundancy in your codebook and waste a lot of bits.

With cascading, you can use a 32d by 128entry (like 32kbit to send the book, 
8bits/frame to send the codeword) fully populated book to whiten the
distribution of the LSPs. Then you can make another pass with a lattice book
to efficently encode the remainder. 

For residues, you can make multiple passes, each time adding more quality.
This makes bitrate pealing work. (without cascading all you can do to reduce
bitrate is totally squash partations, or replace longer codewords with
shorter ones (after the fact bookpareing) which won't have as good of a
perceptual effect..

The basic codebook support for cascading is pretty much in the code aready,
but it's suffered bit-rot while Monty has focused on more basic quality issues.
(block switching, psy model, etc).

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to 'vorbis-request at xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is needed.
Unsubscribe messages sent to the list will be ignored/filtered.