Musical instrument source separation in unison and monaural mixtures

Abstract

Musical Instrument Source Separation aims to separate the individual instruments from a mixture. We work on mixtures where there are two instruments, playing the same constant pitch at the same time. When musical instruments play the same note, they overlap with each other and act as a single source. We use statistical source separation algorithms which perform separation by maximizing the statistical independence between the sources. A mixture can be recorded with more than one microphone; this enables us to extract spatial information of the instruments by comparing the recordings. However we work with the monaural case where there is one microphone. Some musical instruments have amplitude modulation and this modulation can be seen in the mixtures. We also aim to detect amplitude modulations to support the source separation success. We use NTF (Non-negative tensor factorization) to perform the separation. NTF separates the mixture into many components. These components should be clustered in order to synthesize the individual sources. We use k-means as well as manual clustering by comparing the SDR (Signal to Distortion Ratio) values of the components with the original sources.