Drexel University Home Pagewww.drexel.edu DREXEL UNIVERSITY LIBRARIES HOMEPAGE >>
iDEA DREXEL ARCHIVES >>

iDEA: Drexel E-repository and Archives > Drexel Theses and Dissertations > Drexel Theses and Dissertations > Musical ensemble classification using universal background model adaptation and the million song dataset

Please use this identifier to cite or link to this item: http://hdl.handle.net/1860/3883

Title: Musical ensemble classification using universal background model adaptation and the million song dataset
Authors: Dolhansky, Brian
Keywords: Electrical engineering;Sound--Recording and reproducing--Digital techniques;Digital electronics
Issue Date: Jun-2012
Abstract: Obtaining music has become easier with the introduction of digital audio formats and the internet, but the sheer amount of data available makes finding new songs that fit a particular taste a difficult task. Content recommendation services specifically tailored to music have become popular, but most do not analyze audio content, and instead use some form of collaborative filtering which relies on previously-generated metadata. Some music information retrieval (MIR) researchers have instead used content analysis to determine the pertinent attributes of a song, such as the instruments present or the genre. However, the task of determining the musical ensemble that produced a song has not seen much focus. This operation is especially important for song recommendation systems, as users often search for songs produced by a specific ensemble (e.g. a jazz quartet or an a cappella group). In this thesis, musical ensemble classification is covered in depth. Specifically, the novel use of universal background model (UBM) adaptation and the benefits of incorporating the Million Song Dataset (MSD) are examined. It was found that for instrumentation classification, UBM adaptation used in conjunction with the MSD outperformed other traditional machine learning classifiers. These results show that UBM adaptation is a valid method for MIR tasks, and that publicly available unlabeled music datasets can be used to augment the performance of content-based classifiers.
Description: Thesis (M.S., Electrical engineering)--Drexel University, 2012.
URI: http://hdl.handle.net/1860/3883
Appears in Collections:Drexel Theses and Dissertations

Files in This Item:

File Description SizeFormat
Dolhansky_Brian.pdf3.76 MBAdobe PDFView/Open
View Statistics

Items in iDEA are protected by copyright, with all rights reserved, unless otherwise indicated.

 

Valid XHTML 1.0! iDEA Software Copyright © 2002-2010  Duraspace - Feedback