# Aliasing Reduction in Neural Amp Modeling by Smoothing Activations

**Authors:** Ryota Sato, Julius O. Smith  
**arXiv:** [2505.04082](https://arxiv.org/abs/2505.04082)  
**Date:** 2025-05-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.04082v1/

## Key points

- Growing demand for high-fidelity emulations of analog audio hardware, especially vintage guitar amplifiers
- Research focus on neural-network-based black-box modeling techniques in digital audio processing
- Challenge of aliasing artifacts caused by nonlinear activation functions in neural networks
- Study by Ryota Sato and Julius O. Smith on novel activation functions to mitigate aliasing in neural amplifier models
- Introduction of Aliasing-to-Signal Ratio (ASR) metric for accurate evaluation of aliasing extent
- Findings show smoother curve activation functions lead to lower ASR values, reducing aliasing artifacts without significantly increasing Error-to-Signal Ratio (ESR)
- Potential for achieving high modeling accuracy while minimizing aliasing distortions in neural amp models
- Contribution to enhancing digital audio processing techniques and improving reproductions of analog audio hardware through innovative activation function design

## Summary

In the field of digital audio processing, there is a growing demand for high-fidelity emulations of analog audio hardware, particularly vintage guitar amplifiers. This has led to a surge of research into neural-network-based black-box modeling techniques. While deep learning architectures like WaveNet have shown promise in this domain, they still face a persistent challenge - the presence of aliasing artifacts caused by nonlinear activation functions within neural networks. To address this issue, Ryota Sato and Julius O. Smith conducted a recent study exploring novel and modified activation functions designed specifically to mitigate aliasing in neural amplifier models. They also introduced a new metric called the Aliasing-to-Signal Ratio (ASR) to quantitatively evaluate the extent of aliasing with heightened accuracy. In addition to traditional performance metrics such as the Error-to-Signal Ratio (ESR), the study examined a range of established and contemporary activation functions with varying stretch factors. The key findings from their investigation revealed that activation functions characterized by smoother curves tend to yield lower ASR values, indicating a tangible reduction in aliasing artifacts. Importantly, this improvement in aliasing mitigation does not come at the cost of significantly increased ESR values. This underscores the potential for achieving high modeling accuracy while simultaneously minimizing aliasing distortions in neural amp models. This research contributes valuable insights to the ongoing quest for enhancing digital audio processing techniques and offers a pathway towards more faithful and artifact-free reproductions of analog audio hardware through innovative approaches to activation function design. Accepted for presentation at DAFx 2025, this study represents a significant step forward in advancing the state-of-the-art in neural amplifier modeling within the realm of digital signal processing.