AES Europe 2026

This work addresses the problem of frame drum (bendir)
stroke technique recognition in simulated real-world
conditions. The traditional frame drum technique includes
three discrete strokes that are used to create rhythmic
patterns, dum, tek; slap. In the presented work, audio
data augmentation is investigated on a dataset containing
recordings of instruments of various construction
attributes. The used techniques are selected in the
direction of generalizing classification in real-world
conditions. Moreover, the mixing of the frame drum samples
with accompanying guitar chords is introduced, simulating
the more complicated problem of hit technique recognition
when playing in a duo. The application of the
aforementioned data augmentation leads to the formation of
different available datasets for training; testing. Two
convolutional neural network architectures (one-;
two-dimensional) are taken into consideration, trained on
waveforms; melscale spectrograms of the different
subsets accordingly.