I Reverse Engineered the Helix Stadium and Built a NAM-to-Proxy Converter

I’ve built a converter that takes a NAM capture and produces a Helix Stadium Proxy clone. Getting there involved reverse engineering considerably more of the Stadium than I’d planned when I started poking at its Wi-Fi editor last December.

I’ve already had a whinge about wanting NAM on the Stadium. I gig it, I’ve built my presets around it, and the effects and routing are a large part of why it stays on my floor. I’d like to use the captures I enjoy with the effects I already know.

The original project was much smaller. Back in December, I started working out how the official editor talks to the Stadium over Wi-Fi. I watched what happened when I changed things in the app, then built my own software to make the same changes. Eventually I could control blocks and rename snapshots without using the Line 6 editor.

There’s something very satisfying about writing a little bit of code and seeing a piece of guitar gear respond. Particularly when it’s sitting there on the floor looking like it has no interest in your programming hobby.

That work gave me a reason to keep looking. Once I started examining Proxy, I wanted to understand the clone itself: what the Stadium expects from it, and how that becomes the sound coming out. Sending an editor command was one problem. Reproducing the behaviour of an amp model was going to take a lot more work.

Proxy is Line 6’s cloning system. It captures an amp or pedal at a particular setting, then lets you use the resulting clone in a Stadium preset. Neural Amp Modeler also models amps and pedals, with an open-source project behind it. I wanted a way to take an existing NAM capture and make a Proxy version, with the conversion happening on my computer.

Understanding the file format let me create clones, but I also needed to know how Proxy runs the model inside them. I could get the packaging right and still end up with the wrong audio.

I checked my reconstruction against Line 6’s own implementation. Comparing their output sample by sample helped me find mistakes, and eventually the remaining differences were tiny rounding errors. With that working, I could spend my time improving the conversion instead of wondering whether I’d misunderstood how Proxy plays a model.

For the conversion, I use the NAM capture as the reference sound for a new Proxy-compatible model. A capture already contains the amp’s learned behaviour, so I could work from that without getting the original physical amp back out and capturing it again. Getting the new model to follow it closely took considerably longer.

High-gain captures were where I spent a lot of my time. I play that stuff, and the Archon and MT15 models gave me something useful to work on. Distorted guitar gives the converter plenty to get wrong at the start of a note, before you’ve even got to the way it dies away. I wanted the quieter playing to hold up too.

I tried several ways of improving the conversions. Some barely moved the numbers. A few improved one measurement while making another worse. I also tried correcting the output after the model had done its work, which sounded sensible as an idea but didn’t earn its place in the converter when I tested it.

That part gets tedious. You make a change, run the comparison, and discover you’ve mostly found a more elaborate way to get the same answer. More training time wasn’t an automatic improvement either.

The latest refinement did help. On a fresh three-and-a-half-minute guitar recording, average measured error fell by about 21% for the Archon and 12% for the MT15 compared with the previous refined versions. I’d kept that recording out of training, so this was a useful check against different playing. The more difficult sections improved too.

Those gains came from better training within the same model size. Given how much I’ve complained about wanting to leave room for effects in my Stadium presets, I was pleased to get an improvement without making the model bigger.

I’ve kept the audio comparisons and regression tests as part of the project. When I change something, I can compare it with the previous version and check that the exported clone contains the model I intended. Having that machinery in place has saved a lot of guesswork.

I’d still love a NAM block in a Line 6 firmware update. Until then, I’ve got more captures to put through this converter. The Archon and MT15 have had plenty of attention, and I’m curious how the rest of my collection will do.