Linear mixing models for active listening of music productions in realistic studio conditions

The mixing/demixing of audio signals as addressed in the signal processing literature (the "source separation" problem) and the music production in studio remain quite separated worlds. Scientific audio scene analysis rather focuses on "natural" mixtures and most often uses linear (convolutive) models of point sources placed in the same acoustic space. In contrast, the sound engineer can mix musical signals of very different nature and belonging to different acoustic spaces, and exploits many audio effects including non-linear processes. In the present paper we discuss these differences within the strongly emerging framework of active music listening, which is precisely at the crossroads of these two worlds: it consists in giving to the listener the ability to manipulate the different musical sources while listening to a musical piece. We propose a model that allows the description of a general studio mixing process as a linear stationary process of "generalized source image signals" considered as individual tracks. Such a model can be used to allow the recovery of the isolated tracks while preserving the professional sound quality of the mixture. A simple addition of these recovered tracks enables the end-user to recover the full-quality stereo mix, while these tracks can also be used for, e.g., basic remix / karaoke / soloing and re-orchestration applications.