Content Based Record Label Classification for Electronic Music

Characterizing record labels and their sound signatures is a difficult
problem, especially when it comes to indie record labels. That's the
present when dealing with record labels of electronic music, so the
application of different constructs and techniques that have been
proved useful when looking at record labels is essential. On top
of that Electronic music increases the challenges in similar issues,
as timbre and rhythm are more important than root keys, chords,
which are facets traditionally emphasized in music information re-
trieval. The research presented in this dissertation aims at the ex-
ploration of the usage of Music Information Retrieval tools and tech-
niques for the analysis of Electronic Music Record Label based on
audio content. For that purpose we have curated a music collection
especially addressed for the above-mentioned problems, containing
more than 3000 tracks of 9 different Electronic Music Record Label.
The collection has been analyzed with the help of Essentia's library,
and the extracted features present various musical criteria such as
timbre, rhythm, and tonality. The extracted features are tested
with different classication algorithms, from the simplest of them,
a Support Vector Machine, to a Fully Connected layer network, to
a more complex one, a Convolutional Neural Network. We also
propose various ways of segmenting Electronic Music, in order to
capture the most relevant features. We tried to cover as much types
of segments as we can, which were tested experimentally, achieving
satisfactory results. Finally, a detailed qualitative analysis of the
results obtained when considering a group of record labels is per-
formed, demonstrating the potential of the analysis that have been
developed.