Evaluation of a non-negative matrix factorization algorithm for polyphonic guitar to MIDI in real-time

Sammanfattning: This master's thesis is an evaluation of an algorithm for detection and separation of polyphonic guitar notes in real-time. The method used is known as Non-negative Matrix Factorization (NMF) and the algorithm is evaluated for future use in a musical recording and live performance system. The purpose is to let guitarists use the MIDI protocol to play synthesized sounds live using only an electric guitar, a common laptop and a sound card. Requirements for a system like this includes very low latency and high accuracy, demands that work against each other, often leading to a compromise between the two. Implementations of the algorithm are developed in Python and C++ and evaluated with respect to latency, accuracy and feasibility for use on a common laptop. The results shows that the algorithm works fairly well, but the time it takes to gather enough samples for the Fast Fourier Transform to produce results with high enough frequency resolution leads to a too high latency for the purpose of the system. The goal of the thesis is to answer the question whether NMF is a good method for implementation of polyphonic note detection of chords in the system previously mentioned. Based on the results and analysis performed, the thesis concludes in an answer of MAYBE due to yet unsolved problems. Motivations for this answer is provided in the discussion and conclusion sections at the end of the document.