Real-time audio processing with a cascade of discrete-time delay line based reservoir computers

Background: Real-time processing of audio or audio-like signals is a promising research topic for the field of machine learning, with many potential applications in music and communications. We present a cascaded delay line reservoir computer capable of real-time audio processing on standard computing equipment, aimed at black-box system identification of nonlinear audio systems. The cascaded reservoir blocks use two-pole filtered virtual neurons to match their timescales to that of the target signals. The reservoir blocks receive both the global input signal and the target estimate from the previous block (local input). The units in the cascade are trained in a successive manner on a single input output training pair, such that a successively better approximation of the target is reached. A cascade of 5 dual-input reservoir blocks of 100 neurons each is trained to mimic the distortion of a measured guitar amplifier. This cascade outperforms both a single delay reservoir having the same total number of neurons as well as a cascade with only single-input blocks. We show that the presented structure is a viable platform for real-time audio applications on present-day computing hardware. A benefit of this structure is that it works directly from the audio samples as input, avoiding computationally intensive preprocessing.