Abstract | Systems that manage audio databases for sound description and retrieval could be
really useful within a context of music production. In this thesis, one of these systems
has been created using classification techniques, by the application of learning
algorithms and audio feature selections on two datasets. One of them is composed
by commercial sounds and the other one is formed with sounds from an online
repository of Creative Commons audio content. Due to the fact that drums are a
fundamental element on most of the musical genres nowadays, it is the family of
instruments chosen to train and test the presented classification models. Research
is focused on finding generalist drum instrument class and category models, understanding
category as the source nature of the sample (acoustic or digital). Generalization
on these models lead us to be able to classify different drum sound datasets,
achieving good model and prediction accuracies. Combining these models with
an annotation of our datasets with specific values of audio high-level descriptors
(Brightness, Hardness, Roughness and Depth), a drum samples retrieval tool could
be created and would open new possibilities for database management within a music
production framework.
|