Building A Folk Music Recommendation System with LLMs
summary
Shows a concrete pipeline: embed 46,000 folk tunes with OpenAI text-embedding-3-small, cluster (KMeans and variants), visualize with UMAP, and validate clusters by listening to live violin demos — clusters map to jig/reel/waltz and keys. Builds a cosine-similarity vibe search, trains a gradient-boost regressor (reported r² ≈ 0.14) to predict popularity, and iterates an LLM-based ABC tune generator that eventually produces an improved sample (generated tune scored 71).
AI-written summary · about
chapters (13)
- 0:00Introduction to folk music and AI
- 1:09Topic clusters from political manifestos
- 5:26Naming clusters with generative AI
- 8:11Visualizing embeddings with UMAP
- 10:42Folk music data from The Session
- 13:48Embedding and clustering folk tunes
- 17:00Folk tune types with biscuit metaphors
- 19:07Exploring embedding space with audio
- 25:17Interactive folk music discovery
- 27:23Most similar folk tune to Super Mario
- 30:57Predicting folk tune popularity
- 33:45Generating folk tunes with LLMs
- 37:40Q&A
related talks
38mmusic2vec: A smart music recommendation and discovery
35mDiversity and Data: Cases in the Music Industry
41mMusic and Machine Learning
30mLLM-Based Recommendation Systems
43mBuilding your first recommender
51mThe Merger of Machine Learning, Music and Programming
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
45mWorkshop: Building a Music Recommender System in Java