Audio classification of violin bowing techniques: An aid for beginners

Abstract

Playing violin requires both left and right hands that move into one another to produce one distinctive sound. While some violin players improve their hearing and recognize these techniques, it can be difficult for some people. Although there are names and categories for each violin technique, distinctions sometimes become ambiguous. This paper presents an audio classification model utilizing Convolutional Neural Network (CNN) that determines the sound produced by violin and classifies the used technique. The dataset used was gathered from real violin players who were tasked to record themselves playing one specific technique. The recorded tracks were then carefully trimmed to remove the noise. The pre-processed recordings served as an input to a benchmark CNN model. To fully optimize the CNN model, we modified the architecture of the model and tweaked the hyper-parameters. A comparative analysis between the two models was discussed in the latter part of this paper. The result of the analysis showed that our proposed model with an average of 94.8 % accuracy outperformed the benchmark model with an average of 87.6% accuracy. Using stratified cross-validation of five folds, we were able to measure the accuracy, training time, and predicting time of the models. A paired t-test with a p-value of 0.01 that shows a significance between the performance of the two models.

Keywords

• Violin
• Techniques
• Convolutional Neural Network
• Audio classification