Abstract
In this paper, the authors use statistical models to predict the difficulty of recognizing musical keys from polyphonic audio signals. The key recognition difficulty provides important background information when comparing the performance of audio key finding algorithms that often evaluated using different private data sets. Given an audio recording, represented as extracted acoustic features, the authors applied multiple linear regression and proportional odds model to predict the difficulty level of the recording, annotated by three musicians as an integer on a 5-point Likert scale. The authors evaluated the predictions by using root mean square error, Pearson correlation coefficient, exact accuracy, and adjacent accuracy. The authors also discussed issues such as differences found between the musicians' annotations and the consistency of those annotations. To identify potential causes to the perceived difficulty for the individual musicians, the authors applied decision tree-based filtering with bagging. By using weighted naïve Bayes, the authors examined the effectiveness of each identified feature via a classification task.
