A Bottom Up Approach To Audio Source Separation
Keywords
No Thumbnail Available
Authors
Issue Date
2018-07-02
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Humans are e ortlessly able to focus their attention to a speci c sound source in noisy
settings. In a wider sense, single instruments can be seen as a sound source in a song. The
human brain can easily extract the melody of the guitar or the drum pattern out of the song,
just by having the frequency split of the cochlea as input. This paper will investigate how
well deep neural networks (DNN) are able to perform the source separation task. Therefore,
the music theory will be reviewed, in order to have a realistic model of instruments. Then,
current approaches to the source separation problem will be examined, with respect to their
generalization capabilities on di erent data distributions over the train and test set. In the
end, the network will get constrained by instrument theory and biological ndings. The results
were promising for my small models, resulting in a proposition of a di erent convolutional layer
for feature extraction at musical source separation.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen