Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Slimmable NAM: Neural Amp Models with adjustable runtime computational cost

Steven Atkinson

TL;DR

Problem: NAMs for virtual-analog guitar gear are often CPU-intensive, limiting real-time use on CPU-limited devices. Approach: introduce Slimmable NAMs that post-training reduce width from $c$ to $c'$ by truncating $\mathbf{W} \in \mathbb{R}^{c \times c \times k}$ and $\boldsymbol{b} \in \mathbb{R}^{c}$ to $\mathbf{W}' \in \mathbb{R}^{c' \times c' \times k}$ and $\boldsymbol{b}' \in \mathbb{R}^{c'}$, with $d_x=d_y=1$ for mono audio, trained with dry/wet pairs and random $1 \le c' \le c$ per batch, and deployed in a real-time plugin with a GUI slider controlling width $c'$, using a single WaveNet module. Contributions: includes width-truncation method for NAMs, randomized-width training protocol, real-time plugin demonstration, and open-source code. Impact: enables musicians to trade off accuracy and compute on the fly without retraining, broadening accessibility of neural virtual-analog effects.

Abstract

This work demonstrates "slimmable Neural Amp Models", whose size and computational cost can be changed without additional training and with negligible computational overhead, enabling musicians to easily trade off between the accuracy and compute of the models they are using. The method's performance is quantified against commonly-used baselines, and a real-time demonstration of the model in an audio effect plug-in is developed.

Slimmable NAM: Neural Amp Models with adjustable runtime computational cost

TL;DR

Problem: NAMs for virtual-analog guitar gear are often CPU-intensive, limiting real-time use on CPU-limited devices. Approach: introduce Slimmable NAMs that post-training reduce width from to by truncating and to and , with for mono audio, trained with dry/wet pairs and random per batch, and deployed in a real-time plugin with a GUI slider controlling width , using a single WaveNet module. Contributions: includes width-truncation method for NAMs, randomized-width training protocol, real-time plugin demonstration, and open-source code. Impact: enables musicians to trade off accuracy and compute on the fly without retraining, broadening accessibility of neural virtual-analog effects.

Abstract

This work demonstrates "slimmable Neural Amp Models", whose size and computational cost can be changed without additional training and with negligible computational overhead, enabling musicians to easily trade off between the accuracy and compute of the models they are using. The method's performance is quantified against commonly-used baselines, and a real-time demonstration of the model in an audio effect plug-in is developed.

Paper Structure

This paper contains 4 sections, 1 figure.

Table of Contents

  1. Introduction
  2. Method
  3. Results
  4. Conclusions

Figures (1)

  • Figure 1: \ref{['fig:metrics']}: Real-time factor (higher is better) and accuracy (error-signal ratio wright2020real; lower is better) for different models of the four tones considered. The Pareto front is drawn for the slimmable NAM. "New (full)" refers to the a model using the slimmable architecture trained only for use at full size as normal. \ref{['fig:plugin']}: Plugin settings page, showing the ability to slim the loaded network on the fly.