How the pedal plays NAM and IRs

The Dual Stomp Lite can load a Neural Amp Modeler (.nam) amp model into a user AMP slot and a cab impulse response (.wav) into a user IR slot. But the pedal can’t run either one as it is. What it actually runs is much simpler, so every import is an approximation. How good that approximation is depends on the converter.

What the pedal runs instead of a NAM model

A NAM model is a neural network (a WaveNet). The pedal runs a fixed structure instead:

input ──► G1 ──► f ──► G2 ──► output
          512-tap     fixed     512-tap
          filter    saturation  filter

Two FIR filters with a fixed saturation curve between them, a Wiener–Hammerstein model. An import is just the two filters: 2048 floats, of which the pedal plays 512 per filter. That was found by re-amping uploads with test patterns in them (docs/05).

The official app’s conversion measures the model twice: a small-signal frequency response, and a sine tone test (fundamental and third harmonic). It turns each measurement into a minimum-phase filter. It’s fast, but it doesn’t check whether the result sounds like the model on the pedal.

open-stomp’s optimised conversion starts from the same measurements, then fits the two filters to the model:

1. Simulate the pedal. y = G2 ⊛ f(G1 ⊛ x), with the saturation curve f fitted from re-amp recordings. The simulator predicts the hardware within about 0.5 dB.
2. Compare against the model on sweeps, tones and pink noise at several levels, with a multi-resolution spectral loss.
3. Adjust the filters (hand-written gradients, Adam, about a minute) until the simulated pedal matches the model as closely as 512 taps allow. Both filters stay smooth and minimum-phase.

Since version 0.5 the loss weights each octave about the same. Before that the low end counted for almost nothing, which Fixing the low end explains.

What the pedal plays of an IR

A user IR upload is 2048 samples, but the pedal plays only the first 384 (8 ms at 48 kHz). Found the same way: impulses at taps 0–383 come back, later ones don’t.

The official app’s conversion takes the start of the file as it is. The pedal then cuts it off at 8 ms. Many cab IRs start after a short delay or ring for longer than that. Those lose most of their sound, and an IR that starts after the cutoff plays as near silence (35–45 dB too quiet).

open-stomp’s optimised conversion designs a 384-tap minimum-phase filter whose response matches the whole IR as closely as possible. It packs the cab’s tone into the part the pedal plays, then sets the level the app meant to give the IR. It takes under a second.

Limits

Some differences no converter can remove:

• Phase. Both conversions are minimum-phase, so they match how the model sounds (its magnitude response), not its waveform.
• Filter length. A cab with a deep, narrow low-end resonance needs longer filters than 384 or 512 taps. The remaining gap at 60–80 Hz on bass cabs comes from this.
• Nonlinearity. The saturation curve is fixed. High-gain models whose distortion changes character with level or frequency can only be approximated.

Where it’s documented

• NAM and IR uploads: the upload protocol, the official algorithms as recovered from the app, and the optimised conversions in detail.
• Results on hardware: the measurements.
• The analysis code: tools/pedalsim (Python: re-amp rig, pedal simulator, PyTorch prototype of the optimiser, scoring).