Automatic analysis of time varying metrical structures in music

Automatic rhythm analysis is an important area of research in Music Information
Retrieval (MIR) as it aims to develop algorithmic models to study the phenomenon
of beat induction, an action intuitive and almost instinctual to human listeners.
Automatic methods to detect rhythmic as well as musical attributes of a musical
piece like tempo, beats, meter are studied widely in the literature. In this thesis,
we explore time varying metrical structures in music from the perspective of
meter inference and tracking. Time varying metrical structures are a part of various
music cultures and genres like classical music, African music, progressive rock
music etc. This thesis aims to infer and track the musical meter of such musical
pieces by proposing extensions to a data driven Bayesian model which simultaneously
infers the tempo, beats and downbeats of a musical piece. It is shown here
that by adapting this method for time varying metrical structures, the model learns
probabilistic relations for transitions between different rhythm patterns and hence,
metrical changes can be detected within a musical piece. Audio files (containing percussion
patterns) and their annotations (containing information about the metrical
structure) are synthesized to validate the approach. Novel evaluation strategies are
proposed for this case, as the present evaluation measures do not incorporate the
metrical inference accuracy for musical pieces with changing metrical structures.