3.4 Multi-perspective audio description: source
localisation and separation
One characteristic of orchestral music concerts compared
to other amplified musical live performances is how sound
is propagated from the performers to the audience. Sound
sources are spread over a large stage area creating an acoustic
image in front of the audience, which is affected then by
the acoustics of the concert hall. A recording setup might
consist principally of a stereo pairmicrophones placed near
the conductor. In a typical setup, however, this stereo track
can be complemented with a number of zenithal microphones
covering specific instrumental sections. These zenithal
tracks are used to find the right balance in the final mastering
mix.
One of the objectives of the project consists in obtaining
the localisation of the active instruments on stage from a
set of recorded tracks. This process shall include means of
providing a source signal separation. In our scenario, we
might take advantage of additional data such as the score or
source positioning informations (e.g. instrument sections).
State of the art methods of source localisation include
beamforming techniques, which take input signals from
sensor arrays. Other specific techniques address the case of
stereo signals [15]. Regarding source separation, state of
the art techniques involve Non-negative Matrix Factorisation
(NMF) and PLCA [16], butmore recent techniques are
also based on signal-models that exploit musical knowledge
[17]. Score-informed techniques such as [18] are specially
relevant in the context of the project.