Feature Comparison for Classification of Kaustinen Fiddle Playing Style from Archived Recordings Using Deep Learning
Henna Tahvanainen, Tuomas Ylönen, Outi Valo · 2024
The largest collection of the Finnish Folk Music Institute contains the recordings of Kaustinen Folk Music festival since 1968. The recordings have been labelled by hand, and oftentimes the labelling is incomplete. One of the labels that the archive would need is “Kaustinen fiddle playing”, according to the element that was inscribed on the UNESCO's Representative List of the Intangible Cultural Heritage of Humanity in 2021. The playing style is characterized by syncopated and accented rhythms, and compared to traditional violin bowing, the bowing direction is changed off-beat. In this paper, we propose a few neural network classifiers to identify whether the playing style is present in the recording, using either spectrograms or audio descriptors as input. The best performance accuracy (97%) is reached with a four-layer feedforward network using Mel-Frequency Cepstral Coefficients, Mel-spectrogram, chroma, and tonnetz.