Re: [Lame-dev] attack detection

[Robert H. <Rob...@gm...>]

Am Montag, 26. Juni 2006 19:49 schrieb Gabriel Bouvigne:
> Robert Hegemann a =E9crit :
> > it seems there is some regression against 3.98a3 due to the following f=
ix:
> > 'attack detection: workaround against "low freq" periodic signals (like=
 trumpet)'
> > Now we get clearly audible distortions encoding our classical "fatboy"
> > testsample. It seems we miss some needed short blocks here.=20
> > Maybe we can tune the fix for trumpet to give fatboy its short blocks b=
ack?
>=20
> Honestly this was just commited in order to have feedback regarding a=20
> quick workaround that I'd like to include in 3.97.
> I have a whole new attack detection code under work for 3.98.
>=20
> I'll have a look at the workaround in order to try tuning the=20
> thresholds, but I do not think that it would be worth spending too much=20
> time on it.

OK.

Things I want to check out later are:
=2D normalizing the spreading function such that the sum s_i(x) =3D 1
  (maybe some problems in low and high frequency area are due to
  the not normalized spreading function? Want to check this.)
=2D experimenting with some other shape than the current triangle like


Ciao Robert