[vorbis] codebook training

[Monty]

> I try to build codebooks for Vorbis at sampling frequencies other than 32/44/48 KHz. 
> As the first step, I started with residual codebooks. I generated training data on a one minute long 44 KHz audio, using Vorbis compiled with the TRAIN_RES define. I think 
> it worked fine, it generated .vqd files for me. 
> Then I compiled ResVQTrain, which complied OK. Then I run it on the previously generated resaux_short.vqd and resaux_long.vqd files, trying different entries, dim and 
> q_quant parameter values. However the result was the same: 
> - the iteration did not converge, 
> -the metric error was high (in the 1-10 range)
> -the produced .vqi file contained some #IND values as biases.
> -when I run VQBuild on the resulting .vqi files, it reported lots of duplicate entries, and removed.
> 
> Can anyone please tell me what could be wrong?

the training tool you're trying to use hasn't been used since the
early beta days before codebooks were latticed.  That tool won't work
with current releases.

look inside the make_residue_books perl script and the .vqs files it
takes for more information.  In general, however, these tools are not
intended for user or casual developer use.  Codebooks still involve
black magic that I don;t want to spend months documenting.

(Also, training books psecifically for various sample rates won't
actually gain much/any advantage).

Monty

--- >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.