Real-Time Black-Box Modeling With Recurrent Neural Networks

Demo page for the paper submitted to DAFX 2019.

This page includes sound samples processed by the proposed recurrent neural network based model. Two nonlinear audio devices were modeled: the Blackstar HT-1 combo guitar amplifier and the Electro-Harmonix Big Muff Pi. Unprocessed input samples are shown, along with those samples after processing by the target device and by each of the best performing RNN models. The headings indicate whether the model used a GRU or LSTM recurrent unit, as well as the unit's hidden size. All RNN models have just a single recurrent layer, followed by a fully connected layer. The models are ordered with the most accurate (and slowest) first, through to the least accurate (and fastest). For validation the RNN models were compared to three WaveNet models, outputs from the most accurate WaveNet are also included on this page.

A real-time implementation of the proposed deep neural network was developed in C++, using the JUCE framework. The running speed of each model was tested on an Apple iMac with a 2.8 GHz Intel Core i5 processor, the model speed is measured in terms of time required to generate 1 second of output. The model speeds varied from 0.097s - 0.41s per second of output.

Video demonstrating the HT-1 model output with the 'ISF' being adjusted continuously