Irish Folk Music Similarity Analysis

Last month, I successfully defended my master’s thesis, “Irish Folk Music Similarity: Comparing Engineered Features with Deep Learning Embeddings”.

In this work, we created a dataset of Irish folk music gathered from The Session containing more than one recording for every tune. Afterwards, we extracted computational representations using two distinct approaches: feature-engineering and deep learning embeddings. Finally, we compared the two approaches using clustering evaluation metrics to measure how well each representation space recognised the same tune from different recordings.

The main takeaway was that even though none of the approaches provided a clear separation between tunes, after listening to some recordings in the representations yielding the best scores, the deep learning embeddings seemed to aptly capture timbre relationships.

The work can be openly accessed here.

I am looking forward to future opportunities to work on audio and music.