The Mismeasure of Music “On Computerized Music Listening and Analysis via Machine Learning”

Abstract The worthwhile application of computers to study music audio recording collections requires such systems to be engineered to have the appropriate sensitivities and knowledge. One approach to designing such systems is the use of machine learning with music recording datasets. We address two widely held assumptions in the engineering of such systems: 1) if a system reproduces all labels of a music recording dataset then it must have learned about music; and 2) if a system is being trained on a music recording dataset then it is being trained on music. In this chapter, we show that these assumptions are not true, and that machine learning with music recording datasets can result in music listening systems that may not be as successful as they appear. We propose several principles to guide the engineering of computerized music listening and analysis via machine learning.