Tune Similarity Search: Embeddings from ABC Notation for Irish Traditional Music

Listening to Irish music, i’m frequently reminded of similarities between tunes. in some cases, they’re probably mashups of different melodies - in others, perhaps they started from a proto-melody and evolved. The example my teacher brought up was Maggie Pickens sounds an awful lot like Charming Lovely Nancy… and that’s never really left my mind.

The Data

TheSession.org has a massive community-contributed database of traditional tunes in ABC notation — a text format where |:GABc dedB|dedB dedB|c2ec B2dB|A2A2 A2:| represents the musical notes. The dataset has around 20,000 jigs and reels, many with multiple “settings” (different transcriptions of the same tune contributed by different players).

The core problem is: how do you represent a melody in a way that captures its shape, independent of what key it’s played in?

The answer is intervals. Instead of encoding the absolute pitch of each note, I encode the movement between consecutive notes. A tune that goes up a step, up a step, down a third has a distinctive contour regardless of whether it starts on D or G. FolkRNN solved this elegantly by transposing everything to C.

I tried multiple encoding strategies, but settled on Quantized-5, which collapses every interval into one of five buckets: big jump down, small step down, same note, small step up, big jump up. This is deliberately coarse - it throws away a lot of detail, but it’s robust to the kind of variation you see between different settings of the same tune. Turns out Quantized-5 is surprisingly good. Throwing away pitch magnitude information forces the model to focus on melodic contour, which suggests it’s an important signal for perceived similarity. Finer-grained encodings pick up more subtle relationships but also more noise.

The vectors were all loaded into an instance of FAISS - because Annoy (my fave) takes up a lot of disk space and i didn’t feel like deleting anything today. FAISS handles the nearest-neighbor search at build time, and the frontend just looks up pre-computed results.

Does It Work?

Evaluation uses those multi-setting tunes as ground truth. Take a tune with two settings, embed the second setting, and check if the first setting appears in the top 10 nearest neighbors. there is some variability in how true this holds - there are some settings that massively diverge and others that are just slight melodic variations that aren’t captured terribly well (think about how my method would work if you took C-E-F-G and extended the C to a quarter note like C–FG). so it’s not a clean test but it’s decent.

the real test is qualitative — browsing the results and checking whether the similar tunes actually sound similar. When I spot checked, it’s somewhat true at least enough to be useful. The system surfaces tunes that share melodic contour, common phrases, or structural similarity.

Here’s a great example, which has a comment on The Session of: “Reminds me a bit of Drowsy Maggie, but they are different tunes.”

yup!