Improving Audio Retrieval through Content and Metadata Categorization

Audio content sharing on online platforms has become increasingly popular. This necessitates development of techniques to better organize and retrieve this data. In this thesis we look to improve audio retrieval through content and metadata categorization in the context of Freesound. For content, we focus on organization through morphological description. In particular, we propose a taxonomy and thresholding-based classification approach for loudness profiles. The approach can be generalized to structure information about the temporal evolution of other sound attributes. To this end, we also discuss our preliminary findings from extension of this methodology to pitch profiles. On the other hand, metadata systematization has been approached through a topic model known as the Latent Dirichlet Allocation (LDA). Herein automatic clustering of tag information is performed to achieve a higher level representation of each audio file in terms of 'topics'. We evaluate our approach for both the tasks through several experiments conducted over two datasets. This thesis finds immediate application in online audio sharing platforms and opens up several interesting future research avenues. Specifically, evaluation indicates that our methods can be immediately applied to improve Freesound's similarity and context-based search. Moreover, we believe our work on content categorization makes it possible to include an advanced content-based search facility in Freesound