End user would use Halion Sonic which is free, you only need Halion full version to build and edit the library. That's how i remember it anyway.

Mapping XG / AWM2 onto SoundFont Extended (SFe 4) — what fits and what doesn't

I went through the SFe 4 spec to see how cleanly the real XG voice set / MUXG2K's AWM2 engine maps onto it. Short version: SFe 4 fixes most of the plumbing problems that made XG awkward in plain SF2, but its synthesis model is still literally SF2.04 (the spec says so outright), so a few things still don't have a home.

What SFe 4 actually changes (vs SF2.04): 64-bit chunks for huge banks, split bank addressing (wBank → byBankMSB + byBankLSB), variable bit depth + OGG/FLAC compression, UTF-8 names, and a redefinable default-modulator list (DMOD). Generators/units/filter model are unchanged.

The clean fits

• []Voice → Preset → Instrument (1:1)
  []Bank MSB / LSB / program → byBankMSB / byBankLSB / preset#. Splits the bank word, maps 1:1 onto XG's CC0/CC32. Biggest win vs SF2 (no bank-128 drum hack).
  []Drum kits (MSB 126/127) → percussion bank-select flags — addressed by bank-select like XG, not a hardcoded bank.
  []Element (up to 4) → Instrument zones, one per element.
  []Sample → shdr, 16-bit PCM (OGG/FLAC optional).
  []Key / vel split → keyRange / velRange.
  []Pan → pan (0.1% units).
  []Base cutoff → initialFilterFc, in absolute cents: cents = 1200 · log2(Hz / 8.176). This is the per-voice anchor.
  []Resonance → initialFilterQ (cB).
  []Amp envelope (DAHDSR) → volume-envelope generators (convert AWM2 rates to timecents).
  []Same-note choke → exclusiveClass (= drum choke groups).
  []Reverb / chorus → reverbEffectsSend / chorusEffectsSend (send level only).

Where SFe got XG right
It adds the XG sound-controller CCs as default modulators, overridable per-bank via DMOD:

• []CC74 → cutoff: SFe default is 6000 absolute cents (~94 cents/step). On real MU hardware I measured 75 cents/step, so override to ~4800 cents to match.
  []CC71 → resonance, CC73 → attack, CC72 → release all have defaults; CC75 → decay you add yourself.
• Heads up: SFe removes SF2's default velocity→cutoff modulator (it flags itself with isng = "SFE 4"), so you add velocity-brightness explicitly.

What still doesn't map (SF2.04 model ceiling)

1. []Filter key-tracking — no generator for "cutoff follows key." AWM2 does this. Workaround: bake it into multiple key-split zones with stepped initialFilterFc. Most tedious part.
   []Separate pitch + filter envelopes — SF2 has one mod-env feeding both. The MU runs independent PEG and FEG. You get one faithfully, or a compromise.
   []Variation / insertion effects — no equivalent. You get a reverb/chorus send level and nothing else; the MU's actual FX algorithms live outside the format.
   []Velocity→cutoff curve — MU uses a piecewise multi-point curve; SF only has fixed linear/concave/convex/switch shapes. Approximate or bake into velocity zones.

So the bank addressing and the XG-shaped default modulators are genuinely well-suited to an XG dump — feels like they had this use case in mind. The limits are SF2.04's synth model, not SFe's container.

Happy to share measured per-voice cutoff anchors (in Hz) for specific voices if anyone's actually building one of these.

https://github.com/spessasus/syxg2sfe

An attempt at this. it's pretty close.

Just listened to descent midis. I prefer the MU90 real mode 😜

well, yes, a mu90 will be better than a fake mu50. 😀

That's not exactly fair. but it's pretty darn close.

that said, a bug was introduced again. hopefully it will be fixed soon.

A note for people on the fence about getting muxg2k.

The linux version supports both x86-64 and aarch64. i have no idea if any arm SBCs can handle it, btu that owdl certainly be an interestign test. a pi-xg2000 would be quite interesting...

Since the osx version is similar in size, i assume it also supports intel macs and apple silicon.

The windows version seems to be x86-64 only.