Jazz Harmony Embeddings is a transformer-based model designed to generate embeddings from jazz chord progressions. The model processes the chord chart of a jazz tune, tokenized as one token per chord event over a condensed changes skeleton, and outputs a single 128-dimensional vector representation. Its training objective is to position harmonically related tunes—such as those that are transpositions, alternate charts, or contrafacts—close together in the resulting vector space.
This model was trained from scratch on approximately 8,000 chord charts using contrastive learning techniques. It is described as a flat contrastive encoder and employs a 3-seed ensemble approach. The tool is implemented in PyTorch and is associated with the field of music information retrieval, aiming to facilitate tasks that require understanding or comparing the harmonic structure of jazz compositions.
Precomputed embeddings for 6,900 jazz standards are made available, providing immediate resources for analysis or further research. The model, along with its code, evaluation harness, and experiment records, is accessible via a linked GitHub repository. 0) license.
evidence_sufficient": true}
Jazz Harmony Embeddings is a Foundation models & chat product. It focuses on providing embeddings for jazz chord progressions to facilitate music information retrieval and analysis. It is built as an open-source project for music information retrieval researchers and developers. Jazz Harmony Embeddings is open source under the Open Source license. The product ships for the web and the command line.
eigenben builds and maintains Jazz Harmony Embeddings, and the product first shipped in 2026. Key capabilities include chord embedding, music analysis, and contrastive learning.
Latest indexed changes and source events
eigenben/jazz-harmony-embeddings verified by the PulseGate indexer
Other apps tracked under the same category.
