> It totally devastates:
> http://users.belgacom.net/gc247244/logsweep.wav
Oh, to be more specific, the effect of the bug is that short blocks are being
used where they should not be. The current trigger decides to use a short
block when the character of the spectrum changes a great deal over a short
period. What it sees in the above sample is a very strong moving peak; at the
end of the sampling period, the spectrum is entirely different! (One very
strong peak is replaced by a different very strong peak).
Short blocks, in the current encoder setup, are tuned to encode impulses better
than tones. So, because the entire sample is encoded in short blocks, it
sounds like total shit.
Just providing a bit of detail so you know I'm not talking out my ass ;-)
For comparison purposes, here's the same sample with short blocks disabled (and
a less lossy codebook set):
http://www.xiph.org/ogg/vorbis/test.ogg
It's still not perfect (the LSP fit on a single strong frequency is the
pathological worst case right now), but it illustrates what I'm talking about.
There's a third suspected bug in masking at this point that appears to result
in peaks not getting adequate resolution. I'm looking into that one too.
Monty
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
