Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Parametric Neural Amp Modeling with Active Learning

Florian Grötschla, Longxiang Jiao, Luca A. Lanzendörfer, Roger Wattenhofer

TL;DR

This work tackles data efficiency in parametric guitar-amp modeling, where the number of knob settings is $k$ and data collection becomes expensive. PANAMA combines an LSTM-assisted ensemble with a WaveNet-style parametric amplifier and uses gradient-based active learning to select informative knob vectors $\mathbf{g}$ by maximizing the ensemble disagreement $D_{\mathcal{L}}$. The approach collects a small labeled set (e.g., $75$ datapoints) and trains a final feedforward WaveNet that achieves perceptual parity with NAM on MUSHRA tests. It provides an open-source, data-efficient pathway for end-user parametric amp modeling, enabling practical virtual-amp workflows with limited recordings.

Abstract

We introduce Panama, an active learning framework to train parametric guitar amp models end-to-end using a combination of an LSTM model and a WaveNet-like architecture. With \model, one can create a virtual amp by recording samples that are determined through an ensemble-based active learning strategy to minimize the amount of datapoints needed (i.e., amp knob settings). Our strategy uses gradient-based optimization to maximize the disagreement among ensemble models, in order to identify the most informative datapoints. MUSHRA listening tests reveal that, with 75 datapoints, our models are able to match the perceptual quality of NAM, the leading open-source non-parametric amp modeler.

Parametric Neural Amp Modeling with Active Learning

TL;DR

This work tackles data efficiency in parametric guitar-amp modeling, where the number of knob settings is and data collection becomes expensive. PANAMA combines an LSTM-assisted ensemble with a WaveNet-style parametric amplifier and uses gradient-based active learning to select informative knob vectors by maximizing the ensemble disagreement . The approach collects a small labeled set (e.g., datapoints) and trains a final feedforward WaveNet that achieves perceptual parity with NAM on MUSHRA tests. It provides an open-source, data-efficient pathway for end-user parametric amp modeling, enabling practical virtual-amp workflows with limited recordings.

Abstract

We introduce Panama, an active learning framework to train parametric guitar amp models end-to-end using a combination of an LSTM model and a WaveNet-like architecture. With \model, one can create a virtual amp by recording samples that are determined through an ensemble-based active learning strategy to minimize the amount of datapoints needed (i.e., amp knob settings). Our strategy uses gradient-based optimization to maximize the disagreement among ensemble models, in order to identify the most informative datapoints. MUSHRA listening tests reveal that, with 75 datapoints, our models are able to match the perceptual quality of NAM, the leading open-source non-parametric amp modeler.

Paper Structure

This paper contains 13 sections, 4 equations, 4 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Guitar Amp Modeling
  4. Active Learning
  5. Methodology
  6. Model Architecture
  7. Active Learning Approach
  8. Experimental Evaluation
  9. Experimental Evaluation
  10. Active Learning
  11. Ablation against random and heuristic sampling
  12. Ablation on the effect of different architectures
  13. Conclusion

Figures (4)

  • Figure 1: Single Model Setup. The parametric amp model transforms the DI guitar input signal, conditioned on the settings of virtual knobs.
  • Figure 2: Active Learning Setup. The input signal $\mathbf{x}$ and amp settings $\mathbf{g}$ are fed into independently trained instances of the model, and variation in the outputs is calculated as the cross-model disagreement $D$. Gradients are propagated back to $\mathbf{g}$ in order to maximize $D$.
  • Figure 3: Distribution of component values in $\mathbf{g}$-vectors gathered through active learning.
  • Figure 4: Comparison of MUSHRA scores across models. Bars show the mean score and error bars represent the 95% confidence interval. "Ours-10/25/75” indicate our models trained with 10, 25, and 75 active learning samples, respectively. Ours-75 matches the quality of NAM, the leading open-source non-parametric amp modeler.