Back Members of the MTG provide the best solution for the automatic recognition of musical genres

Members of the MTG provide the best solution for the automatic recognition of musical genres

It was presented at the WWW2018 Congress, held from 23 to 27 April in Lyon (France) and won the WWW 2018 Challenge: Learning to Recognize Musical Genre. The study involved Xavier Serra and Minz Won, together with researchers from the Delft University of Technology (the Netherlands).

15.05.2018

 

The automatic recognition of musical genres on the basis of audio information is a challenge facing music technology because often, the tags that describe musical genres attributed manually by humans are considered subjective and noisy (confusing). For example, a rock tune may also be considered pop, or many classical works can also be instrumental. In turn, tagging depends on the context of the listener, for example, a French person would not classify a song in French as “international”. In general, there is no universal genre taxonomy and so classification is a challenge.

This was the starting point of research in the field of automatic recognition presented at the WWW2018 Conference, which was held from 23 to 27 April in Lyon (France) and provided the best solution to the 2018 WWW Challenge: Learning to Recognize Musical Genre for the correct prediction of musical genres of unknown audio segments using the FMA (Free Music Archive) dataset.  

Gràfic del procés seguit pels investigadors.

This automatic recognition procedure based on audio information is set out in an article of which Xavier Serra and Minz Won, director and doctoral researcher, respectively, of the Music Technology Group (MTG) at the Department of Information and Communication Technologies (DTIC) at UPF are co-authors, along with researchers from the Delft University of Technology (Netherlands), Kim Jaehun, first author and Cynthia C.S.  Liem.  

The work sets out from the basis that in the classification of music, the tags that describe the artists are less subjective and less noisy (confusing) than those that describe musical genres and, in turn, some artists are strongly related with certain musical genres. “In this work, we propose applying the transfer learning framework, learning from different types of information related to the artist or artists, and then use it to infer genre classification”. In the challenge, this solution has proved to be the most efficient and brings greater solidity to the tag, minimizing the potential noise”, claim the authors of the study. Six teams took part in the challenge and in two rounds they had to send their genre classification predictions for 35,000 clips of 30 seconds each. The team made up of Minz Won and Kim Jaehun, was the winner.

Reference work:

Jaehun Kim, Minz Won, Xavier Serra, Cynthia C.S. Liem (2018), “Transfer Learning of Artist Group Factors to Musical Genre Classification”, WWW ‘18 The Web Conference 2018, 23-27 April, Lyon (France).

Multimedia

Categories:

SDG - Sustainable Development Goals:

Els ODS a la UPF

Contact