Slimmable NAM: Neural Amp Models with adjustable runtime computational cost

AI summary

Overview

Research area: Machine learning for audio, specifically data-driven virtual analog modeling of musical equipment (guitar amplifiers), with a focus on real-time inference efficiency in an audio effect plug-in.

Technical level: Intermediate. The method itself is simple (weight truncation), but the context assumes familiarity with WaveNet-style convolutional architectures, audio model-training workflows, and real-time audio performance metrics.

One-sentence scope: The paper introduces "slimmable" Neural Amp Models whose width — and therefore runtime computational cost — can be reduced after training, with negligible overhead and no additional training, and demonstrates the approach against baselines and in a real-time plug-in.

What This Paper Is About

Neural Amp Modeler (NAM) models are trained by other people and then used by musicians, who cannot easily retrain them to fit their own CPU constraints. The usual workaround, model distillation, requires ML compute such as a GPU and disrupts creative workflows. This paper asks whether a NAM can be made to shrink on demand — like truncating a finite impulse response (IR) — so musicians can trade accuracy for compute on the fly.

Key Contributions

1. A slimming procedure for WaveNet-based NAMs. Convolutional weights $\mathbf{W} \in \mathbb{R}^{c \times c \times k}$ and biases $\boldsymbol{b} \in \mathbb{R}^{c}$ are truncated to $\mathbf{W}^{\prime} \in \mathbb{R}^{c^{\prime} \times c^{\prime} \times k}$ and $\boldsymbol{b}^{\prime} \in \mathbb{R}^{c^{\prime}}$, reducing network width $c$ to $c^{\prime}$. For the affine input and output projections (input dimension $d_x$, output dimension $d_y$, with $d_x = d_y = 1$ for mono audio), only the rows of the input projection and the columns of the output projection are truncated, so input and output dimensionality is preserved.

2. A training scheme that makes slimming free. During training, each mini-batch randomly slims the network to a chosen $1 \leq c^{\prime} \leq c$ and supervises the slimmed network's predictions against the targets, so the resulting network works at any number of channels without retraining.

3. Open-source training and real-time prediction code plus an installable plug-in (referenced as Atkinson 2025a, 2025b, 2025c), with a GUI slider that adjusts $c^{\prime}$ so users can audition different model sizes in real time.

4. A quantified comparison against commonly-used baselines on four real amplifier tones, using a new single-WaveNet architecture rather than NAM's standard stacked two-WaveNet design (as of release v0.12.0).

Main Findings

• Slimming costs essentially nothing at runtime. Model size and computational cost can be changed without additional training and with negligible computational overhead.

• One trained model covers a range of sizes. Because $c^{\prime}$ is sampled uniformly ($1 \leq c^{\prime} \leq c$) during training, a single trained network can be used with any number of channels at prediction time.

• Evaluation spans four amplifier tones. A "clean" tone from a Fender Deluxe Reverb, a "crunch" tone from a Morgan MVP23, and hi-gain "rhythm" and "lead" tones from an Omega Ampworks Obsidian.

• The trade-off is presented with two metrics. Real-time factor (higher is better) and accuracy measured as error-signal ratio (Wright et al. 2020; lower is better). Figure 1(a) plots these for the slimmable NAM alongside other commonly-used models, with a Pareto front drawn for the slimmable NAM.

• A useful control condition is included. "New (full)" refers to a model using the slimmable architecture trained only for use at full size, serving as a point of comparison against the slimmable version.

• Specific numerical values are not reported in the paper text. The quantitative results appear only in Figure 1(a) as plotted curves; the paper text does not state numerical real-time factors or error-signal ratios.

• A working real-time demonstration exists. Figure 1(b) shows the plug-in settings page, adapted from NeuralAmpModelerPlugin, where moving the slider adjusts $c^{\prime}$ to audition models of different sizes on the fly. Audio examples and model file links are provided via a video demonstration.

Methodology in Plain English

A WaveNet is a stack of convolutional layers operating on a time series of $c$-dimensional vectors. The approach is to make those vectors narrower. To slim a network from width $c$ to width $c^{\prime}$, you literally keep a smaller slice of each layer's weight tensor and bias vector — the same kind of operation a musician might do when truncating an IR. The input and output projections are handled asymmetrically (rows for the input, columns for the output) so that the model still accepts one audio channel in and produces one audio channel out, matching mono audio where $d_x = d_y = 1$.

The training trick is what makes this work without extra training runs. Ordinary training would only ever optimize the full-width network, and truncating it afterward would degrade badly. Instead, every mini-batch picks a random width between 1 and the full width $c$, slims the network to that width, and compares its output to the target. Over training, the network learns to perform acceptably across the whole range of widths. NAM's usual training procedure applies: supervised learning on a "dry/wet" pair of audio files. The authors then built the slimming control into the settings page of the plug-in so the user can move the slider and hear the trade-off immediately.

Why This Matters

Impact on research: The paper connects NAM to the broader literature on slimmable networks (Yu et al. 2019), dynamic gating and conditional computation (Ehteshami Bejnordi and Krestel 2020), multi-exit and early-exit networks (Saeed 2022), and supernets (Cai et al. 2020). Slimmable networks have already been applied in speech (Elminshawi et al. 2025), and pruning has been explored for virtual analog modeling (Südholt et al. 2022) but requires additional training. This work's distinction is placing the compute/accuracy trade-off in the hands of the end user rather than the model-maker.

Real-world applications:

• Live performance on limited hardware. A guitarist whose laptop or pedal-board processor cannot run a full NAM model can slim it down in the plug-in until it runs reliably, rather than seeking a lighter model elsewhere.
• Recording sessions with many simultaneous tracks. Slimming individual amp models reduces the aggregate CPU load of a project without swapping models or committing to a lower-quality render.
• Consumer audio effect plug-ins. The slider exposes a model-size control directly in the settings page, turning a technical constraint into a user-facing tone/resource knob.
• Model distribution. A single model file with a variable-size architecture could replace shipping several separately trained sizes of the same tone.

Industry relevance: The slimmable WaveNet architecture introduced here is under investigation as the next-generation default model architecture for NAM, with the stated hope of increasing the project's accessibility to musicians. Since NAM has seen widespread adoption and its users normally do not train their own models, an architecture that removes the compute barrier without complicating the model-fitting process is attractive to both plug-in developers and the musicians who play the models.

Future Directions

• Adoption as NAM's default architecture. The slimmable WaveNet is under investigation as the next-generation default model architecture for NAM; whether it is adopted, and how it interacts with existing model files, remains open.
• Extending slimming to NAM's other architectures. The paper focuses on the stacked WaveNet due to its strong performance but states the method can be adapted to NAM's other open-source architectures in a similar way — that adaptation is not demonstrated here.
• Building out the stacked architecture. This demonstration deliberately uses a new single-WaveNet module instead of NAM's standard stacked two-WaveNet design for simplicity; how slimming behaves in the stacked configuration is not reported.
• Quantifying and characterizing the trade-off more fully. The reported results are plotted in Figure 1(a) with no numerical values in the text, leaving room for a finer-grained analysis of accuracy degradation as a function of $c^{\prime}$ across tones and for comparison against distillation and pruning under matched budgets.

Target Audience

• Musicians and producers who use NAM or similar neural amp models and are constrained by CPU in real-time settings.
• Audio plug-in developers interested in exposing compute/accuracy controls directly to users.
• Machine learning researchers working on audio who want a concrete, low-complexity example of width-slimmable training applied to virtual analog modeling.
• Practitioners in model compression looking for a case study contrasting slimming (no retraining) with distillation and pruning (additional training), and for the connection to supernets and early-exit networks.

Authors’ 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.