Source separation in single-channel musical mixtures
- Members: Mr Georgios Siamantas, Dr John Szymanski
- Start Date: June 2008
Audio and, in particular, musical signal separation is a highly interdisciplinary area of research, with a wide variety of applications. By musical source separation we mean the identification and extraction of the signals that coexist in an audio mixture and correspond to the sounds coming from a number of musical sources. We are focusing on cases where there is only one observation of the mixture available (i.e. single-channel mixtures). Since the amount of initial information in these cases is very limited compared to multi-channel approaches, additional information is required to assist the separation. This information is going to be chosen and employed in such a way so that (1) our approach can be applied to a wide variety of mixtures and (2) it does not require excessive user intervention or prior learning/training procedures. This separation approach could be useful in applications such as the following: Music information retrieval tasks (automatic music transcription, classification of musical material, melody/artist similarity, etc.), creative audio effects, musicology and restoration of old recordings.