ImprovNet - Generating Controllable Musical Improvisations with Iterative Corruption Refinement

Keshav Bhandari 1 , Sungkyun Chang 1 , Tongyu Lu 2 , Fareza R. Enus 1 , Louis B. Bradshaw 1 , Dorien Herremans 2 , Simon Colton 1 · 2025

Read ImprovNet - Generating Controllable Musical Improvisations with Iterative Corruption Refinement by Keshav Bhandari 1 , Sungkyun Chang 1 , Tongyu Lu 2 , Fareza R. Enus 1 , Louis B. Bradshaw 1 , Dorien Herremans 2 , Simon Colton 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-04522.txt.

Description

Keshav Bhandari and six co-authors present ImprovNet, a transformer model that improvises on existing music in a chosen style. Performance-level style transfer for whole pieces is hard, and jazz data are scarce. Their model learns by a corruption-refinement game: it practices repairing deliberately damaged music, which teaches it how musical elements fit together.

One model then covers many jobs: restyling a piece within or across genres, harmonizing a melody in a genre's voice, continuing a short prompt, and filling gaps. Users control how far the output drifts from the original. In listening tests, 79 percent of participants correctly heard classical pieces converted to jazz. The model beat a leading music transformer at continuation and infilling tasks. The authors release code and a demo.

Catalog record

CREATORKeshav Bhandari 1 , Sungkyun Chang 1 , Tongyu Lu 2 , Fareza R. Enus 1 , Louis B. Bradshaw 1 , Dorien Herremans 2 , Simon Colton 1
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-04522

Browse the full library · About the Tower of Babel