Retro-gaming circles have argued for decades about which General MIDI module plays DOS game music “correctly.” John Novak settled it with recordings instead of opinions: 46 classic DOS soundtracks, rendered on 7 different MIDI modules — 322 lossless recordings, all published on the Internet Archive along with the MIDI files and REAPER projects.

The method is the point. The same captured MIDI stream was fed to every module, so the synthesizer is the only variable:

  • Games run in DOSBox Staging, MIDI output captured through loopMIDI into REAPER at 120 BPM / 960 PPQ
  • Hardware (Roland SC-55, Yamaha MU80) recorded at 24-bit/48kHz through the DAC of a Yamaha MG10XU mixer, with output knobs set by ear to avoid distorting the analog stage
  • Softsynths rendered offline, faster than real time
  • Output loudness-matched and time-aligned, with a REAPER project holding one track per module for honest A/B switching

What the 322 recordings actually show:

  • SC-55 vs Roland Sound Canvas VA — the forum claim that the software version emulates badly is mostly noise. The SCVA is a recreation of the later SC-8820, and it sounds close enough for everyday gaming; what it misses is the subtle analog saturation that gives the hardware its tighter bottom end.
  • SC-55 vs Yamaha MU80 — the Yamaha wins 12–9 on his listen-through and is objectively the better device: lower noise floor, no polyphony limits, better samples and effects. The usual counter-argument (SC-55-composed music falls apart elsewhere) shows up far less often than the internet claims.
  • Later Sound Canvas models — close to useless for DOS games. Duke Nukem 3D (1996), Azrael’s Tear (1996) and Stonekeep (1995) are all confirmed composed on the SC-55, so the “everyone had upgraded to an SC-88” theory has little evidence behind it.
  • Hardware vs softsynth — Yamaha’s free S-YXG50 VSTi is close at sample level; the real MU80 has a high-end shimmer and mid scoop the plugin lacks.

The honest part is the ending. After a week of switched listening, Novak admits he would be hard-pressed to identify these modules in a blind test, and that if you own just one you simply never notice anything wrong. These were $500+ devices that most people never compared side by side — the search for the “correct” module is a collector phenomenon.

Still, if he had to pick one: hardware Yamaha MU80. And the whole comparison is reproducible, because the data is published rather than described.