Xarxes neuronals per a la generació i representació d'ones

Abstract

During the last few years, neural networks have become a relevant discipline in the field of machine learning. In this project, several neural networks architectures were proposed for the representation and generation of sound waves. The final goal of this project is the generation of sounds to accompany silent videos, reprodu- cing the Foley Efect. However, it was required to study how to obtain the sound characteristics and how to generate audios. To determine if the arquitectures are suitable, two different data sets were used: a group of synthetic notes, and a set of sounds extracted from tennis matches. In the representation of waves, autoenco- ders networks were used to obtain a feature vector. Two representation types were employed: the sound wave, and the espectrogram. To generate waves, sound was created using Google Wavenet network. The amplitud was predicted as a function of the previous amplitudes.