Convolutional neural network for chordophones recognition
Nicholas Dutz, Nick Gangemi, Andrew Cunningham, Michael Kvartunas, Amelia Vignola, Joseph F. Vignola, Hannah R. Kurdila · The Journal of the Acoustical Society of America · 2019
A student project at Catholic University related to string instruments compared the radiated sound from a guitar, a banjo and 10 ukuleles using a convolutional neural network, CNN. The objective was to determine if the CNN could be trained to distinguish between the instruments and classify them. The guitar, banjo and one of the ukuleles were factory built and the remaining 9 ukuleles where built by the team, four from kits and five from scratch. There were at least 100 single note strikes recorded for each instrument and each instrument was recorded in three different acoustic environments, anechoic, highly reverberant and moderate reverberation. This presentation will discuss the construction process used to build the ukuleles and test method used for recording. Additionally, data comparing the scratch build instruments to comparably sized factory built ukuleles will be presented to show that the CUA built instruments were appropriate for evaluation. Finally, the presentation will show the network’s predictions for instrument type and recording environment.