Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Anti-aliasing of neural distortion effects via model fine tuning

Alistair Carson, Alec Wright, Stefan Bilbao

TL;DR

Neural distortion models often suffer aliasing when nonlinearity generates harmonics beyond the Nyquist limit. The paper introduces a teacher–student fine-tuning framework where a frozen teacher produces an alias-free target $\tilde{y}_{\rm teach}$ by removing non-harmonic components from $y_{\rm teach}=f(x, \theta_{\rm teach})$, and a student $y_{\rm stud}=f(x, \theta_{\rm stud})$ learns to reproduce this target by minimizing $\mathcal{L}=\mathrm{ESR}(y_{\rm stud}, \tilde{y}_{\rm teach})+\lambda\mathrm{NMR}(S_{\rm stud}, \tilde{S}_{\rm teach})$ with $\lambda=1$. The method applies a pre-emphasis low-pass filter and trains on synthetic sine tones to create aliasing-free supervision, enabling anti-aliasing in both open-weight LSTMs and TCNs as well as custom models trained on analog pedals. Objective results show significant aliasing reduction, often outperforming 2× oversampling, though harmonic content can shift depending on the model, with LSTMs typically providing the best balance between anti-aliasing and preserving analog-like harmonics. Perceptual tests indicate improved similarity to analog references for sine sweeps after fine-tuning, while guitar/bass signals may retain highest similarity when using LSTMs; overall, the approach offers a practical way to mitigate aliasing without runtime inefficiency. The work highlights a viable path to robust neural distortion models with improved perceptual fidelity and broad applicability to neural audio effects.

Abstract

Neural networks have become ubiquitous with guitar distortion effects modelling in recent years. Despite their ability to yield perceptually convincing models, they are susceptible to frequency aliasing when driven by high frequency and high gain inputs. Nonlinear activation functions create both the desired harmonic distortion and unwanted aliasing distortion as the bandwidth of the signal is expanded beyond the Nyquist frequency. Here, we present a method for reducing aliasing in neural models via a teacher-student fine tuning approach, where the teacher is a pre-trained model with its weights frozen, and the student is a copy of this with learnable parameters. The student is fine-tuned against an aliasing-free dataset generated by passing sinusoids through the original model and removing non-harmonic components from the output spectra. Our results show that this method significantly suppresses aliasing for both long-short-term-memory networks (LSTM) and temporal convolutional networks (TCN). In the majority of our case studies, the reduction in aliasing was greater than that achieved by two times oversampling. One side-effect of the proposed method is that harmonic distortion components are also affected. This adverse effect was found to be model-dependent, with the LSTM models giving the best balance between anti-aliasing and preserving the perceived similarity to an analog reference device.

Anti-aliasing of neural distortion effects via model fine tuning

TL;DR

Neural distortion models often suffer aliasing when nonlinearity generates harmonics beyond the Nyquist limit. The paper introduces a teacher–student fine-tuning framework where a frozen teacher produces an alias-free target by removing non-harmonic components from , and a student learns to reproduce this target by minimizing with . The method applies a pre-emphasis low-pass filter and trains on synthetic sine tones to create aliasing-free supervision, enabling anti-aliasing in both open-weight LSTMs and TCNs as well as custom models trained on analog pedals. Objective results show significant aliasing reduction, often outperforming 2× oversampling, though harmonic content can shift depending on the model, with LSTMs typically providing the best balance between anti-aliasing and preserving analog-like harmonics. Perceptual tests indicate improved similarity to analog references for sine sweeps after fine-tuning, while guitar/bass signals may retain highest similarity when using LSTMs; overall, the approach offers a practical way to mitigate aliasing without runtime inefficiency. The work highlights a viable path to robust neural distortion models with improved perceptual fidelity and broad applicability to neural audio effects.

Abstract

Neural networks have become ubiquitous with guitar distortion effects modelling in recent years. Despite their ability to yield perceptually convincing models, they are susceptible to frequency aliasing when driven by high frequency and high gain inputs. Nonlinear activation functions create both the desired harmonic distortion and unwanted aliasing distortion as the bandwidth of the signal is expanded beyond the Nyquist frequency. Here, we present a method for reducing aliasing in neural models via a teacher-student fine tuning approach, where the teacher is a pre-trained model with its weights frozen, and the student is a copy of this with learnable parameters. The student is fine-tuned against an aliasing-free dataset generated by passing sinusoids through the original model and removing non-harmonic components from the output spectra. Our results show that this method significantly suppresses aliasing for both long-short-term-memory networks (LSTM) and temporal convolutional networks (TCN). In the majority of our case studies, the reduction in aliasing was greater than that achieved by two times oversampling. One side-effect of the proposed method is that harmonic distortion components are also affected. This adverse effect was found to be model-dependent, with the LSTM models giving the best balance between anti-aliasing and preserving the perceived similarity to an analog reference device.
Paper Structure (18 sections, 17 equations, 6 figures, 1 table)

This paper contains 18 sections, 17 equations, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Input sine tones
  4. Aliasing removal
  5. Pre-emphasis filtering
  6. Loss function
  7. Training details
  8. Validation dataset and metrics
  9. Experiments
  10. Open-weight models
  11. Custom models
  12. Objective results
  13. Custom pre-training results
  14. Fine-tuning results -- aliasing reduction
  15. Fine tuning results -- harmonic analysis
  16. ...and 3 more sections

Figures (6)

  • Figure 1: Fine-tuning procedure for anti-aliasing of the Student model. The dashed line indicates the flow of gradients to the Student parameters. Spectral plots are included for illustration only -- training operates in the time domain (except for the NMR calculation).
  • Figure 2: Output magnitude spectra of the Broadcast (a) and JHM8 (b) Teacher models (i-ii), their respective Student models (iii-iv) and the reference (v) for an input tone of 2394.3. Crosses mark the harmonic components, and (vi) shows the error in magnitude of these w.r.t. the reference.
  • Figure 3: Output magnitude spectra of the Teacher models (top), the corresponding Student models (middle) and the relative error in magnitude of the harmonic components (bottom) for the open-weight Goat (a), Mesa (b), Vox (c) and JCM (d) models. The input tone had $f_0$ = 2394.3 and amplitude -6.
  • Figure 4: NMR-S (i) and HESR-R (ii) against input sinusoidal frequency, $f_0$, for the Broadcast (a) and JHM8 (b) models. The input gain was -6. The black dashed line at -10 indicates the approximate threshold of aliasing audibility Lehtonen_2012_saw. Lower values are better.
  • Figure 5: Response to a sine sweep for the Broadcast LSTM models: (a) Teacher (b) Teacher 2x oversampled (c) Student (d) Student 2x oversampled. The minimum amplitude visible is -80.
  • ...and 1 more figures