From Ludo2026 to Your Studio: Exploring 93,000+ Sega Genesis FM Presets

In August 2026 we packed a year of notebook work into twenty minutes and took it to Ludo2026, the Fifteenth European Conference on Game Music and Sound, held at Edge Hill University. Our talk, Exploring the Sonic Universe of the Sega Genesis: Data Science, FM Synthesis, and 93,000+ Presets, sat in Session 9, Synthesis, Served Two Ways (see the full Ludo2026 programme), alongside Joseph Chang’s histogram work on 8-bit percussion. This post is the friendly maker’s cut of that deck: what we showed, what the data said, and how you can explore the same corpus yourself.

We are Abraham Casas and Maria Cerro of Kasser Synths, handmade Yamaha FM modules from Santander, Spain. Day to day we bring three chip worlds back into musicians’ hands: the YM2151 of the arcades, the YM2612 / YM3438 of the Mega Drive / Genesis, and the YMF262 (OPL3) of 1990s Sound Blaster PCs. At Ludo2026 we stayed with the Genesis: six channels of four-operator FM, plus the TI SN76489 PSG kept on board for Master System compatibility.

Why treat presets as a corpus?

Ludomusicology already listens carefully to scores, drivers and cultural context. What is still underexplored is the mountain of voice recipes sitting inside VGM logs: the ALG, feedback, TL and envelopes that composers wrote into the chip while the game ran. A VGM file is not a rendered WAV; it is a sample-accurate command stream of register writes. From that stream the community has long extracted portable OPM banks. The seed of our study is the dump DrWashington posted on KVR in February 2010: Project2612 VGMs batch-run through Shiru’s vgm2opm. We first told that discovery story in Part 1: Discovery.

Our pipeline turns those raw OPM files into an analysis-ready dataset: parse each voice (~58 chip parameters), deduplicate by timbre (ignoring TL-only volume clones), normalise game names, and enrich with composer, nationality and GEMS usage from Sega Retro, VGMrips and Wikipedia. The cleaned corpus holds 93,000+ presets. The notebooks, methodology and web app are open source on GitHub.

What the study shows

We never ask the data for genre labels. We only ask which presets look alike in YM2612 parameter space. Still, a few clear patterns emerge.

• Shared YM2612 grammar. Across composers and regions, Algorithm 4 (two modulated carriers summed) and maximum feedback (FL = 7) dominate as practical defaults. Feedback itself behaves almost binary in the corpus: FL = 0 or FL = 7, with little middle ground.
• Feature scores that travel. We derived a Brightness Index and a Complexity Score (both 0–1) so 58 knobs become discussable. Brightness clusters mid–high (~0.55–0.7), enough bite for small TVs and dense mixes, while Complexity sits upper–mid (~0.6–0.85).
• Games leave fingerprints. Pulseman commits hard to Algorithm 5 (~75% of its presets). Gunstar Heroes, Streets of Rage 3 and Thunder Force IV spread across algorithms: identity from variety rather than from a single wiring.
• Tools matter. Sega of America’s GEMS driver is a minority in our corpus (about 7.3% of presets, 43 of 547 games) but not a footnote. GEMS leans Algorithm 2 and FL = 0: industrial conditions and multi-track reliability, not “worse FM”.
• Composer neighbourhoods sit inside the toolbox. Credits for Yuzo Koshiro (Streets of Rage), Masato Nakamura (Sonic), Norio Hanzawa (Gunstar Heroes) and Naoki Kodaka (Super Fantasy Zone) show recognisable habits (envelopes, routing outliers, house palettes) without rejecting the chip defaults. For more on that Japanese soundscape, see our composers post.

Those findings are hypotheses grounded in parameter space. A preset is not a full score (melody, sequencing and PCM drums live elsewhere), and today’s clustering does not yet hear the audio. Future work we sketched at Ludo2026 is exactly that: render each patch, extract audio features, and recommend by timbre as well as by knobs.

Seven sonic realms and a live map

The unsupervised path is simple to say and careful to run: compress with PCA, group with K-Means (elbow at K = 7), lay out with UMAP for a clickable 2D map. Only chip parameters enter the fit: no composer, game or genre labels. Human names come later, by listening. We call the seven emergent groups sonic realms:

• Raw Signals
• Neon Action
• Polished Arcade
• High-Speed Chiptune
• Deep Space FM
• Fantasy Atmospheres
• Experimental Playgrounds

That map is DAFMExplorer: filter by realm, game or composer, play presets through a software YM2612 in real time, and save up to six voices as a DefleMask DMP bank (here is how to load DefleMask DMP presets on real hardware). We told the full project story (notebooks, embeddings, and why we built it) in Exploring the Sonic Universe: When FM Synthesis Meets Data Science. The live app is here:

https://dafm-explorer.vercel.app/ (best on desktop or tablet landscape).

An invitation, not a closed paper

We brought DAFMExplorer to Ludo2026 because presets are evidence of shared timbral practice on a platform: measurable, still audible, openly reusable. The conference theme was “Pushing limits and crossing borders”; for us that border sits between listening and counting, between chiptune nostalgia and reproducible notebooks, between a Beamer slide and a Eurorack panel with a real Yamaha die.

If you want the long-form methodology, start with the DAFMExplorer post and the discovery origin story. If you want the chip under your fingers, the DAFM Genesis is waiting. And if you want to fork the corpus, the code is on GitHub: change the clustering, join another archive, or take the same pipeline to YM2151 and YMF262 next. We will keep mapping. The rest of the sonic universe is still out there.

Take the DAFMExplorer presets from Ludo2026 to the real YM2612!

Prefer to solder it yourself? Get the DAFM synth GENESIS DIY kit on Tindie. Also at Perfect Circuit.