Time-Varying Audio Effect Modeling By End-to-End Adversarial Training

Yann Bourdin^1,2, Pierrick Legrand^2,3, Fanny Roche¹

¹Arturia, F-38330 Montbonnot-Saint-Martin, France

²Inria center at the University of Bordeaux, Astral Inria Team, F-33405 Talence, France

³IMS, UMR CNRS 5218, ENSC, Bordeaux INP, F-33405 Talence, France

Paper

Source code repository

Slow-LFO Spectrograms

Fast-LFO Spectrograms

Detailed results for Experiment 2

Below are some highlighted models for which we present output spectrograms using a chirp-train signal, and musical examples that can be listened to. All our results are available through the links above.

Slow-LFO

Fast-LFO

① Adversarial Training	② Fine-tuning	① Adversarial Training	② Fine-tuning
λ=0.01 ε=0.01 Figure 8c Chirp-train Clean Electric Guitar Electric Guitar House Loop	λ=0.01 ε=0.01 Chirp-train Clean Electric Guitar Electric Guitar House Loop	λ=0 Figure 8j Chirp-train Clean Electric Guitar Electric Guitar House Loop	λ=0 Figure 9l Chirp-train Clean Electric Guitar Electric Guitar House Loop
λ=0.1 ε=0.01 Figure 8f Chirp-train Clean Electric Guitar Electric Guitar House Loop	λ=0.1 ε=0.01 Figure 9j Chirp-train Clean Electric Guitar Electric Guitar House Loop	λ=0 Chirp-train Clean Electric Guitar Electric Guitar House Loop	λ=0 Chirp-train Clean Electric Guitar Electric Guitar House Loop
λ=0 Figure 8h Chirp-train Clean Electric Guitar Electric Guitar House Loop	λ=0 Figure 9i Chirp-train Clean Electric Guitar Electric Guitar House Loop	λ=0.01 ε=0.1 Chirp-train Clean Electric Guitar Electric Guitar House Loop	λ=0.01 ε=0.1 Figure 9o Chirp-train Clean Electric Guitar Electric Guitar House Loop

Reference (Target Effect's Output)


Chirp-train Clean Electric Guitar Electric Guitar House Loop	Chirp-train Clean Electric Guitar Electric Guitar House Loop