Music Audio Signal Prediction using Machine Learning

Abstract

Even though considerable advancements have been made in time series forecasting for audio, there are still many unexplored aspects. An objective of the analysis is to develop a viable product to replace the look-ahead functions of audio dynamic range compressors. Towards this end, and given the suitability of neural networks for predictive purposes, this project discusses the application of MultyLayer Perceptrons (MLPs) and Long-Short Term Memory (LSTMs) for addressing this research question. The numerical experiments focuses on the predictions of this systems. It is analyzed how changing window length (number of inputs), prediction steps (number of outputs), and sampling frequency (dataset resolution) affects prediction quality. The findings indicate that, after a threshold, increasing number of inputs yields diminishing rewards.