[vorbis-dev] Argh, preecho mistriggers

[Gregory Maxwell]

On Fri, 16 Jun 2000, Monty wrote:

> In case people on the list don't know it, A Vorbis tester (who's email I've 
> lost!  If you're on the list, speak up!) is maintaining a vorbis audio quality 
> test page:
> 
> http://hem.passagen.se/ingets1/vorbis.htm
> 
> He's doing us the great service of hunting down samples that break Vorbis (so
> we don't have to :-)
> 
> In any case, it turns out that the trumpets in the dr4 sample are reliably mistriggering the preecho detector (which implies either a large bug in the detector, or general borkkenness which must be fixed).  The artifacts in the sample are due entirely to Vorbis trying to encode the trumpets entirely in short blocks (which will suck for strongly periodic/harmonic waveforms, as produced by brass instruments).
> 
> Just a heads up; if someone else runs across this artifact (it manifests as
> 'wayyy to much noise/rasp'), you know that I knwo about it. 

By the way, do we use the LSP hysterisis on short as well as long
blocks? If not, that might improve the situation.. (Actually, looking
breifly at the hysterisis code, we *are* keeping the states seperate
between short and long right?)

I ask because it sounds like bad LSP fit due to the shortblocks not being
trained well for strongly tonal signals. 

Another hysterisis suggestion: If some criteria is not met (like the last
short block was more then 3 blocks ago, or there is some sort of other
detectable change) then perform a brute force lookup. The slowdown from
doing that every block would be too much, but perhaps, the slowdown from
doing it every once in a while might be acceptible.. It might improve
quality by giving a better inital match for the hysterisis..
 

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/