Paper written by Nicolas Brochec*, Tsubasa Tanaka*, and Will Howie*† has been presented at the ICMC2024 (International Computer Music Conference 2024).
* Tokyo University of the Arts
† Japan Society for the Promotion of Science
Abstract: Within existing research on the automatic classification of musical instrument playing techniques, few available datasets include enough playing techniques to cover the full range of a given musical instrument’s expressive ability. However, creating a new large dataset requires recording many samples for many performance techniques, which is costly and time-consuming. Therefore, in this study, we attempt to augment data by increasing the number of recording microphones without increasing the recording duration and verify the effectiveness of this data augmentation method. As a result of recording flute playing techniques using multiple microphones, the accuracy and macro F1-Score of a convolutional neural network-based classifier improved when using a combination of the five most close-to-source microphones. The classifier’s performance further improved when data were combined with a data augmentation method based on pitch shifting.
