Learning musical instruments from mixtures of audio with weak labels

David Little*, Bryan A Pardo

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contribution

17 Scopus citations

Abstract

We are interested in developing a system that learns to recognize individual sound sources in an auditory scene where multiple sources may be occurring simultaneously. We focus here on sound source recognition in music audio mixtures. Many researchers have made progress by using isolated training examples or very strongly labeled training data. We consider an alternative approach: the learner is presented with a variety of weaky-labeled mixtures. Positive examples include the target instrument at some point in a mixture of sounds, and negative examples are mixtures that do not contain the target. We show that it not only possible to learn from weakly-labeled mixtures of instruments, but that it works significantly better (78% correct labeling compared to 55%) than learning from isolated examples when the task is identification of an instrument in novel mixtures.

Original languageEnglish (US)
Title of host publicationISMIR 2008 - 9th International Conference on Music Information Retrieval
Pages127-132
Number of pages6
StatePublished - Dec 1 2008
Event9th International Conference on Music Information Retrieval, ISMIR 2008 - Philadelphia, PA, United States
Duration: Sep 14 2008Sep 18 2008

Other

Other9th International Conference on Music Information Retrieval, ISMIR 2008
CountryUnited States
CityPhiladelphia, PA
Period9/14/089/18/08

ASJC Scopus subject areas

  • Music
  • Information Systems

Fingerprint Dive into the research topics of 'Learning musical instruments from mixtures of audio with weak labels'. Together they form a unique fingerprint.

Cite this