Machine Learning Based Prediction of MIDI Signals for Networked Music Performance Applications

Networked Music Performance (NMP) is envisioned as a potential game changer among Internet applications: it aims at revolutionising the traditional concept of musical interaction by enabling remote musicians to interact and perform together through a telecommunication network. Ensuring realistic performative conditions, however, constitutes a significant engineering challenge due to the extremely strict requirements in terms of audio quality and, most importantly, network delay. Unfortunately, such requirements are rarely met in today’s Internet. When an audio signal is streamed from a source to a destination, audio data is divided into packets; the delivery of packets to the destination is subject to an unavoidable delay due to transmission and propagation over the physical medium, plus a variable jitter. Therefore, some or even most packets may not reach the destination in time for the playback, which causes gaps in the audio stream to be reproduced. This thesis proposes the adoption of machine learning techniques to conceal missing packets carrying MIDI audio signals, by predicting future events that will be generated by the MIDI source. Results show that the proposed approaches based on feedforward and recurrent artificial neural networks outperform a baseline model whose task is repeating the last played notes given as input.