Module optimise 

Optimised conversion: fit the two filters so the pedal itself sounds like the model (docs/05).

The official converter designs G1 and G2 from two measurements and never looks at what the pedal does with them. This module simulates the pedal instead and adjusts G1/G2 until the simulated output matches the model’s on a test signal:

• The pedal (fitted from re-amp recordings, accurate to the rig’s own repeatability): y = G2 ⊛ f(G1 ⊛ x), with both filters cut to PEDAL_TAPS and f the fixed saturation curve saturation.
• The target: the model’s output for the same signal (input −12 dB, DC-blocked, scaled to the pedal’s output level).
• The filters: smooth log-magnitude curves at 41 control points (20 Hz × 2^(i/4)), each made minimum-phase, as the official converter’s filters are.
• The fit: Adam on a multi-resolution STFT log-magnitude distance over sweeps, tones and pink noise at several levels, starting from the Measure::Clean conversion. Gradients are written out by hand (no ML framework).

A port of tools/pedalsim (PyTorch); tests/data/optim holds its reference values.

Structs

Optimised

The result of optimise.

Options

Optimiser settings (the defaults are the ones validated on hardware).

Segment

A named part of the TestSignal.

Simulator

The pedal simulator for one input signal (its block spectra are computed once).

TestSignal

The signal the optimiser matches the pedal to the model on (all at SR).

Enums

Weighting

How the loss weights the STFT bins.

Constants

CURVE_POINTS

Control points of the filter curves: 41, a quarter octave apart from 20 Hz (to 20.48 kHz).

SR

Sample rate of the test signal and the simulation.

Functions

optimise

Fit G1/G2 so the simulated pedal matches model, starting from the responses in init (2048 floats of upload data, normally a Measure::Clean conversion). Runs the model over the test signal once, then options.steps simulations; about a minute per 300 steps in an optimised build.

optimise_with_progress

optimise, calling on_step(done, total) after each step (the model render before the first takes about a sixth of the time). Returning false stops early with the best result so far.

render_target

What the pedal should output for x: the model’s output (input −12 dB, from its reset state), through the app’s DC blocker, at the pedal’s output level.

saturation

The pedal’s fixed saturation curve f, applied between G1 and G2 (piecewise linear).

simulate

The pedal’s output for x with user AMP data (2048 floats as uploaded).

test_signal

The test signal (about 53 s): the re-amp signal the pedal model was fitted on (log sweeps at −36/−24/−12/−6 dBFS, bass tones at three levels, pink noise at −18 dBFS) followed by pink noise at −36, −24 and −12 dBFS. The fit uses everything but the −18 dBFS pink noise.