Procedural generation of sound files from small sound samples

Abstract

Algorithmic composition is a method by which individuals can generate sounds without human intervention, through the use of some sort of algorithm. It is an area of research that has gained a lot of popularity within the last 20 years because of the ever-growing market for games, movies and TV which make use of it. Its applications also extend to non-developers, whereby anyone can take their favorite rain sound file and algorithmically compose longer sequences of it. In this project, we designed two audio synthesisers to determine which of the two produce the most euphonious and natural sounding results. The two machine learning methods that were used for the development of the audio synthesisers were Recurrent Neural Networks and Markov Chains. Investigation was also carried into various Recurrent Neural Network approaches for the procedural audio generation to determine which method will yield the most euphonious sounds. So as to streamline the process and allow for automatic tagging of all inputted files, an audio classifier using a Convolutional Neural Network was also built. Through this classifier inputs are interpreted, processed and grouped so as to be used by the synthesiser to procedurally generate audio. Through the use of a Convolutional Neural Network, our audio classifier was able to label our inputs with over 70% accuracy. This made the process of sequencing our audio all the easier. The final audio sequencer was also developed to a satisfactory stage. Audio can be sequenced through the use of both the Markov Chain and the Recurrent Neural Network. By use of a well developed survey, we also determined which of these systems produced the most natural sounding results. We elected to divide our results section into two areas, natural sounds and non-natural sounds. Examples of natural sounds being thunder and rain, and non-natural sounds being cars, engines and human speech. This was done so as to better analyze the results. Results showed overwhelming favour towards Markov chains in natural sounds. Multiple reasons are at play for this result, however the leading cause was that the Recurrent Neural Network tended to output repetitive sequences when producing sequences for natural sounds. The results for non-natural sounds show there is a split vote and that the system used to develop sounds can be interchanged without much difference being noted in the resultant quality of the outputs.