Re: [fluid-dev] FluidSynth problem

[Josh Green]

On Tue, 2003-12-02 at 03:03, Joerg Anders wrote:
> Hi all!
> 
> This is my first mail on this list. Perhaps some of you are
> on the linux audio users list and you already know the problem.
> 
> FluidSynth is a good software. But is has problems with certain
> patches. For instance patch 49 in GM (strings).
> 

It would be good to give FluidSynth a good comparison workout between
other synths. I did a little bit of testing of that nature myself, but
hardly scientific :)

The numbering in FluidSynth starts with Preset 0, so just to clarify do
you mean "Strings" which is number 48 (FluidSynth program #)?

> Please have a look at:
> 
> http://rnvs.informatik.tu-chemnitz.de/DROP/gap.html
> 
> At the end of the page is an example file: strings.mid
> 
> It consists of 2 tied whole note chords played by 
> strings (patch 49)
> 
> You'll find there the sound as MP3 rendered by:
> 
>       - hardware midi synthesizer:
>       - TiMidity
>       - FluidSynth
> 
> I used the (copyrighted) 8mbgmsfx.sf2 soundfont
> (which is shipped with some soundcards).

> Below you'll find the same with the free FluidR3
> soundfont. (from: www.hammersound.net)
> 
> In both cases FluidSynth creates a wrong sound.
> In contrast to all other sythesizers it creates
> "attacking" strings. 
> 
> Unfortunately, the string rendering is very important.
> Please have a look at:
> 
> http://rnvs.informatik.tu-chemnitz.de/SYNTH/synth.html
> 
> Here you'll find a piece of music played by a bigband.
> Again all is renderd using the same soundfont.
> 
> But the FluidSynth rendering is practically unuseable.
> 
> Does anybody know a solution? Some users on
> linux audio user list found out that this patch
> (49) consists of 8 layers. Perhaps this is the reason why.(?)

8 layers? In the 8MBGMSFX.SF2 I have there is actually 9, but they
aren't overlapping or anything so only one layer should sound per note.
Perhaps you might be able to figure out whats causing the problem if you
were to use Swami which uses FluidSynth (thats usually what I use when
trying to figure out these types of problems). Of note is that Swami
does its own instrument handling, and loads these instruments into
FluidSynth. So there is the possibility of difference between the two
implementations. Having the patch loaded in Swami might give you a
better idea of what might be causing it, though. You could also listen
to the samples directly to see if perhaps its the sample itself. There
is quite a lot of deviation between different SoundFont implementations,
FluidSynth could still use some tweaking in regards to emulating other
synths. The volume attenuation in particular is a problem with EMU8K
SoundFont files, but thats because FluidSynth follows the SoundFont spec
and EMU8k cards don't. The result is that the sounds will be too quiet
with FluidSynth.
I tested out 48 (strings), I don't think I have the best ear for this
type of stuff, but it didn't sound like it was as dramatic an attack as
the FluidSynth sample you put on that web page. Of note though is that
the attack is set to minimum value, so in theory it should be played
back just as the sample full volume. Its nice to see someone doing some
comparisons. I don't have any particular ideas of whats at fault here,
but will let you know if I find anything. Cheers.
        Josh Green