Table of Contents: :: Library Catalog

Saved in:

Bibliographic Details
Main Author:	Chawla, Varun
Format:	Recurso digital
Language:
Published:	Zenodo 2026
Online Access:	https://doi.org/10.5281/zenodo.18811689
Tags:	Add Tag No Tags, Be the first to tag this record!

Table of Contents:

Abstract—Text-to-music models today are trained almost entirely on Western music. They do not understand ragas, talas, or the ornamentation central to Indian traditions, and they produce generic output when prompted for Indian genres. This paper describes RagaLoRA, a LoRA adapter that tunes ACE-Step 1.5’s Diffusion Transformer decoder to generate Indian music across ten genres—Hindustani classical, Carnatic, Bollywood, qawwali, ghazal, bhajan, and four more. The adapter is rank-16, weighs 44MB, touches 11M of the model’s 2.4B parameters (0.46%), and trains on 250 balanced samples drawn from a 2,787-segment corpus of Creative Commons and public domain Indian recordings. In A/B tests against the unmodified base model, the adapter shifts outputs in genre-appropriate directions: classical and devotional genres get darker and slower (spectral centroid drops up to 16%, tempo drops up to 35%), while dance and rock genres get louder (RMS energy rises up to 38%). Five of the ten evaluated genres had zero dedicated training data, yet still show coherent shifts, pointing to transfer across related Indian styles. No prior LoRA adapter for a diffusion music model has targeted Indian music at this breadth. Index Terms—music generation, Indian music, LoRA, diffusion transformer, raga, fine-tuning

Similar Items