The potential for automatic assessment of trumpet tone quality

The goal of this study was to examine the possibility of training machine learning algorithms to differentiate between the performance of good notes and bad notes. Four trumpet players recorded a total of 239 notes from which audio features were extracted. The notes were subjectively graded by five brass players. The resulting dataset was used to train support vector machines with different groupings of ratings. Splitting the data set into two classes (―good‖ and ―bad‖) at the median rating, the classifier showed an average success rate of 72% when training and testing using cross-validation. Splitting the data into three roughly-equal classes (―good,‖ ―medium,‖ and ―bad‖), the classifier correctly identified the class an average of 54% of the time. Even using seven classes, the classifier identified the correct class 46% of the time, which is better than the result expected from chance or from the strategy of picking the most populous class (36%).