Affordable Access

Core Learning

Publication Date
  • Virtues Character Strengths
  • Computer Science
  • Musicology


DR AF T 4TH INTERNATIONAL WORKSHOP ON COGNITIVE INFORMATION PROCESSING, MAY 26–28, 2014, COPENHAGEN, DENMARK A CLOSER LOOK AT DEEP LEARNING NEURAL NETWORKS WITH LOW-LEVEL SPECTRAL PERIODICITY FEATURES Bob L. Sturm, Corey Kereliuk Audio Analysis Lab, AD:MT Aalborg University Copenhagen A.C. Meyers Vænge 15 DK-2450 Copenhagen, Denmark Aggelos Pikrakis Dept. of Informatics University of Piraeus 80 Karaoli & Dimitriou Str. 18534, Piraeus, Greece ABSTRACT Systems built using deep learning neural networks trained on low-level spectral periodicity features (DeSPerF) repro- duced the most “ground truth” labels of those submitted to the MIREX 2013 task, “Audio Latin Genre Classification”. We take a closer look at a DeSPerF system to test if it is learning and using criteria relevant to the task of music genre recog- nition, or whether it is using irrelevant criteria confounded with the “ground truth” labels of a dataset. We find DeSPerF systems can obtain high figures of merit when “ground truth” labels are confounded with low-level temporal factors in the training and testing datasets. 1. INTRODUCTION In the 2013 MIREX “Audio Latin Genre Classification” task (ALGC), systems created using deep learning neural networks of low-level spectral periodicity features (DeS- PerF) [1] reproduced the most “ground truth” labels of the 10-class Latin Music Dataset (LMD) [2], producing a nor- malized classification accuracy of 0.77.1 This is a remarkable result for two reasons. First, music in the classes of LMD are described by its curators [2] as having complex intrinsic and extrinsic characteristics. For example: “[Axe´] originated in Salvador”; “[Bachata is] closely associated with poor rural migrants”; “The basic instruments [of Forro´] are the accor- dion, a triangle and a zabumba”; “The lyrics [in Gau´cha] usually refer to a specific set of values like respect for the women”; and, “The most important aspect of the instrumen- tation in Salsa is the Clave.” Second, DeSPerF uses very lo

There are no comments yet on this publication. Be the first to share your thoughts.


Seen <100 times