[Paper Review] Randomized Overdrive Neural Networks

Christian J. Steinmetz, Joshua D. Reiss|arXiv (Cornell University)|Oct 8, 2020

Neural Networks and Applications13 references4 citations

TL;DR

This paper proposes Randomized Overdrive Neural Networks (RONN), a novel method for generating creative audio distortion effects by using randomly initialized Temporal Convolutional Networks (TCNs) in the time domain. Without training, architectural parameters like depth, kernel size, activation function, and weight initialization directly shape diverse sonic characteristics—from subtle overdrive to extreme temporal smearing resembling reverb—enabling real-time, interactive sound design via a VST/AU plugin with full control over network architecture.

ABSTRACT

By processing audio signals in the time-domain with randomly weighted temporal convolutional networks (TCNs), we uncover a wide range of novel, yet controllable overdrive effects. We discover that architectural aspects, such as the depth of the network, the kernel size, the number of channels, the activation function, as well as the weight initialization, all have a clear impact on the sonic character of the resultant effect, without the need for training. In practice, these effects range from conventional overdrive and distortion, to more extreme effects, as the receptive field grows, similar to a fusion of distortion, equalization, delay, and reverb. To enable use by musicians and producers, we provide a real-time plugin implementation. This allows users to dynamically design networks, listening to the results in real-time. We provide a demonstration and code at https://csteinmetz1.github.io/ronn.

Motivation & Objective

• To explore the creative potential of untrained, randomly initialized neural networks for audio signal processing.
• To develop a real-time, interactive audio plugin that allows users to design and audition novel distortion effects by adjusting neural network architecture.
• To demonstrate that architectural choices in TCNs—such as depth, kernel size, and activation function—directly influence sonic character without requiring model training.
• To enable musicians and producers with no machine learning expertise to access complex, controllable audio effects through intuitive, dynamic network parameterization.
• To investigate whether fully convolutional, randomly initialized networks can produce compelling, musically useful effects comparable to traditional distortion and spatial processing.

Proposed method

• The method employs a feedforward, fully convolutional Temporal Convolutional Network (TCN) with causal and dilated convolutions to enable large, exponentially growing receptive fields without training.
• Networks are constructed using randomly initialized weights, with no backpropagation or learning, relying solely on architectural hyperparameters to shape the audio output.
• The implementation uses PyTorch's C++ API to embed a parameterized neural network module within a JUCE-based VST/AU plugin for real-time audio processing.
• A look-back buffer is used to maintain signal continuity across block-based processing, ensuring no perceivable discontinuities at frame boundaries.
• Computational efficiency is improved via depthwise separable convolutions, enabling real-time performance even with large receptive fields (up to 4 seconds).
• Stereo output is generated by allowing multiple output channels, enabling cross-channel interactions and spatialized effects from mono or stereo inputs.

Experimental results

Research questions

• RQ1 Can randomly initialized, untrained neural networks generate musically useful and controllable distortion effects in the time domain?
• RQ2 How do architectural components—such as network depth, kernel size, activation function, and weight initialization—impact the sonic character of the resulting audio effects?
• RQ3 To what extent can a fully convolutional, untrained TCN emulate or surpass traditional distortion, delay, and reverb-like effects?
• RQ4 Can a real-time, interactive plugin interface be designed to allow non-experts to explore the space of neural network architectures for audio effect creation?
• RQ5 What is the impact of receptive field size and dilated convolutions on the temporal and spectral characteristics of the processed audio?

Key findings

• Architectural parameters such as depth, kernel size, and activation function directly and controllably shape the sonic character of the output, producing effects ranging from conventional overdrive to extreme temporal smearing resembling reverb.
• Sigmoid activations produced harsh, gritty distortion, while ReLU activations generated fuzz-like effects, demonstrating that nonlinearity shape significantly influences timbre.
• Weight initialization schemes were found to impact the tonal quality of the distortion, indicating that random initialization alone introduces perceptible sonic variation.
• The use of depthwise separable convolutions enabled real-time processing of networks with receptive fields up to 4 seconds on standard CPUs, making complex effects accessible on general-purpose hardware.
• Stereo output generation through multiple output channels enabled cross-channel interactions, producing spatialized and immersive audio effects from mono inputs.
• A global seed control allowed for effect recallability and preset saving, enhancing usability for musicians and producers.