> My first test of vorbis however was a bit disappointing. Testcase
> was gspi*.wav from SQAM. It had a strange sounding tremolo
> effect, but I consider vorbis still alpha. Perhaps something was
> temporarly broken 8�\
Not temporarily broken, but you did get bit by a shortcoming already corrected
on a CVS branch.
Specifically, Vorbis encodes its output using vector quantization; rather than
brute forcing the encoding process (by trying each of several hundred codeword
possibilities looking for closest fit), it uses a Monte-Carlo trained
partitioning tree to find best fit in a few comparisons rather than a few
hundred.
The monte-carlo training was [is] the weak link. The training data had
nothing in it that looked anything like gspi_?.wav, meaning that the decision
tree was completely incapable of modelling this sound. It was added to the
training set and that cured the problem.
Monty
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
