In this thesis we address a fundamental aspect of the computational analysis of Indian
art music, the automatic identication of the tonic of the lead performer. We propose
two approaches for tonic identication in Indian art music, which take advantage of the
characteristic feature of this music tradition by performing a multi-pitch analysis.
We provide a short introduction to Indian art music, explaining the concept of tonic
in the context of this music tradition. We review the main audio features, techniques and
works relevant to the computational analysis of the tonal aspect of music and present a
critique of previous approaches to tonic identication in Indian art music.
A detailed description and the implementation steps for both the proposed methods
are presented. The audio signal is transformed using a multi-pitch representation,
which is then used to construct the pitch histograms. The tonic is identied using the
prominent peaks of a pitch histogram. Following a classication based approach the
system automatically learns the best set of rules to select the peak of the histogram that
represents the tonic. In addition to the multi-pitch representation, the second method
also analyzes the predominant melody pitches to estimate the tonic octave. Further,
we also present a proposal for a complete iterative system for tonic identication which
aims to use both audio and metadata.
The methods are evaluated on a sizable diverse database of Indian art music, compiled
as a part of the CompMusic project. The obtained results are good and demonstrate
the advantage of using a multi-pitch approach. A detailed error analysis is performed
and the plausible reasons for errors are discussed. The thesis is concluded with
a summary of the work, highlighting the main conclusions and and the contributions