Vorbis dev - Jan 2004 - A bitrate peeling attempt

Given all this talk of peeling lately, I thought I'd try out a method
I've been thinking about for a while now. It goes something like this:

* Peeling is achieved by dropping some of the residue.
* The encoder indicates how much of the residue to retain (e.g a number
of stages, 3 bits in the stream per packet, non-standard!).
* The decoder (peeler) copies the contents of packets except for the
residue stages to drop.

I've got the decode-side of peeling working - it's about 10 times faster
than full decoding (e.g oggdec), due to not needing MDCTs or analysis.
It's a bit annoying that it still involves going through most of the
motions of real decoding because of the variable length codes everywhere
(especially the codebooks).

However... I'm a bit stuck trying to get the encoder (libvorbis) to
cooperate. First thing to do was remove that hard coded 3 stage limit
all over the place (easy enough). I then tried modifying an existing
static_bookblock to have the first stages of _resbook_44s_1, followed by
_resbook_44s_3, using 6 stages overall. The idea here is that it would
generate a q1 followed by a q3 residue. But this generates a file with a
huge amount of error.

What exactly should I be doing to generate residues like I describe
above?

- John Ripley.
--- >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-dev-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

Had a bit of progress here...

...
> 
> However... I'm a bit stuck trying to get the encoder (libvorbis) to
> cooperate. First thing to do was remove that hard coded 3 stage limit
> all over the place (easy enough). I then tried modifying an existing
> static_bookblock to have the first stages of _resbook_44s_1, followed by
> _resbook_44s_3, using 6 stages overall. The idea here is that it would
> generate a q1 followed by a q3 residue. But this generates a file with a
> huge amount of error.
...

The garbled sounding output turns out to be either a problem with what
I've done to the encoder, or an actual bug in the decoder. It happens
because some of the VQ codebooks are empty, e.g
"_vq_lengthlist__44c1_s_p2_0" (in res_books_stereo.h). The codes are
therefore zero length, and are written out as... nothing!

So, when the decoder grabbed a code (decode_packed_entry_number), it got
an error and the residue unpack stopped prematurely. A simple test for
an empty codebook before getting a code fixed this and resulted in a
decoded file which sounds the same as the high quality version. Which
was nice.

Trouble is the peeled version sounds completely garbled compared to the
same size low quality encode (I'm just arbitrarily retaining the first 4
stages for now). Guess I need to toy with the template a bit. But it
does work, at least a little :)

- John Ripley
--- >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-dev-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.