Abstract
A system and method determine whether a candidate audio catalog contributed to the training of a generative audio model. An elemental encoder decomposes generated audio and catalog items into elemental embeddings corresponding to aesthetic elements including melodic contour, harmonic progression, rhythmic pattern, and timbral palette. A segmentation engine defines regions of the generated output for scoped analysis. Within each element independently, instance-level scores compare output regions against catalog items while distribution-level alignment compares the catalog's element distributions against distributions over model-generated outputs relative to a reference set. A multi-order analyzer scores recurrence patterns, including motif persistence. A provenance scorer combines all evidence into a provenance score, and an attribution report generator decomposes it into per-element, per-region detail with confidence. A personalization engine tunes a class-conditioned matcher for catalogs too small for whole-distribution testing. Applications include rights-holder assertion, platform compliance auditing, and licensing evidence generation.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Zavesky, Eric, "SYSTEM AND METHOD FOR ELEMENTAL DECOMPOSITION-BASED PROVENANCE ATTRIBUTION OF AI-GENERATED AUDIO", Technical Disclosure Commons, (September 02, 2026)
https://www.tdcommons.org/dpubs_series/11548