Series-Parallel Time-Variant Modulation (SPTVMod)

This repository contains the PyTorch implementation of SPTVMod used for the December 2025 pre-print paper Time-Varying Audio Effect Modeling By End-to-End Adversarial Training authored by Y. Bourdin, P. Legrand and F. Roche.

Link to the paper: https://inria.hal.science/hal-05413743

Accompanying website: https://ybourdin.github.io/sptvmod/

Architecture

(1a)	(1b)

Figure 1. Block diagram of the architecture of SPTVMod. The generator has an orange background, and the discriminator a yellow one.

Figure 2b. Composition of the processing blocks: ModBlock (left), FXBlock (upper right) and FeatBlock (lower right).

Training

• Use the set_target_length() method to compute:
  • the expected input length for the specified target length;
  • the index intervals to slice the tensors given to different parts of the model, here the modulation path and audio path;
  • the cropping sizes of the cropping layers.
• To obtain the generator's output, call forward() with paddingmode = CachedPadding1d.NoPadding, and use_spn = True if using state prediction.
• To obtain the discriminator's outputs, compute the feature lists of the true and predicted outputs with disc_features and use disc_from_features() method to obtain the discriminator's outputs.

Computing the slicing indices and the cropping sizes

This computation is unchanged from SPTMod: https://github.com/ybourdin/sptmod?tab=readme-ov-file#computing-the-slicing-indices-and-the-cropping-sizes

Dataset

The Fast-LFO and Slow-LFO datasets we used for our experiments are available in the folders dp4_fast_lfo_dataset and dp4_slow_lfo_dataset.