Neural Networks and Music Synthesis using Sound Data Sets

Abstract

This article proposes a comparative study between two topological structures of Neural Networks – Recursive Neural Networks (RNN) and WaveNet – applied to sound synthesis and analysis of sound datasets. Based on these two specific systems, the state of the art of these technologies in the field of contemporary sound creation was evaluated so that it could be possible to identify technical limitations and aesthetic possibilities for applying these systems in musical contexts. The relevance of the research in implementing these models in the field of sound creation and the Brazilian context focuses on the critical study of the adequacy of machine learning techniques in synthesis and the aesthetic implications of this technology in composition practices. At the current research stage, we conclude that the application of these synthesis methods falls short of professional use since the sounds produced have a high noise index, have low resolution and hardly maintain compositional coherence over the time of the samples. We also emphasize that implementing these systems in the Brazilian context is problematic, as developing these models requires access to costly high-performance computational resources. Hence, we have identified that a possible alternative to this problem of access to adequate infrastructure is the subscription to processing services via the cloud – we emphasize, however, that they are part of a monopoly of technology companies located exclusively in the Global North.