Pruning Deep Neural Network Models of Guitar Distortion Effects

Deep neural networks have been successfully used in the task of black-box modeling analog audio effects such as distortion. Improving processing speed and memory requirements inference step is desirable to allow models be on a wide range hardware concurrently with other software. In this paper, we propose new application recent advancements network pruning methods recurrent distortion using Long Short-Term Memory architecture. We compare efficacy method four different datasets; one pedal three vacuum tube amplifiers. Iterative magnitude allows us remove over 99% parameters from some without loss accuracy. evaluate real-time performance pruned find that 3x-4x speedup can achieved, compared an unpruned baseline. show training larger model then it outperforms equivalent hidden size. A listening test confirms does not degrade perceived sound quality, but may even slightly improve it. The proposed techniques design computationally efficient deep for electric guitar real time.