Slimmable Neural Amp Modeler Models (arxiv.org)

Researchers introduced "Slimmable NAM" — a family of neural amplifier (amp) models whose effective size and runtime compute can be changed on-the-fly without retraining and with negligible overhead. By adapting slimmable-network ideas to neural amp modeling, a single trained model exposes multiple operating points so musicians and developers can trade off fidelity for latency and CPU use at inference time. The paper quantifies performance against common baselines and demonstrates a real-time audio-effect plug-in, showing the approach is practical for live use. Technically, Slimmable NAMs let you select smaller or larger sub-networks at runtime (adjusting width/complexity) while sharing weights, so there’s no need to train separate models for different resource budgets. Because switching costs are minimal, the method supports adaptive inference on devices with changing compute constraints (e.g., mobile, embedded DSP, or low-latency desktop setups). For the AI/ML community this offers a compelling alternative to separate model compression pipelines (pruning/quantization) for audio tasks: one model can serve multiple deployment points, simplifying development and enabling dynamic, real-time quality/compute scaling for creative audio applications.