Futga: Towards Fine-grained Music Understanding through Temporally-enhanced Generative Augmentation

TL;DR

FUTGA tackles the limitation that existing music captioning methods produce only global, short clips descriptions and ignore temporal structure. It introduces temporally-enhanced generative augmentation to synthesize full-length songs from short MusicCaps clips, annotate them with time-boundary segments, and enrich global captions with an LLM, guided by MIR features. The framework fine-tunes the SALMONN-7B model on synthetic data and uses MIR-feature alignment and human feedback (Harmonixset prompts) to produce accurate, time-aware captions that are then used to augment MusicCaps and Song Describer datasets. Empirical evaluation across caption generation, retrieval, and music generation shows improved fine-grained captions, better temporally-aware retrieval, and beneficial conditioning for text-to-music generation, while highlighting sim-to-real gaps and suggesting directions toward long-context and QA tasks.

Abstract

Existing music captioning methods are limited to generating concise global descriptions of short music clips, which fail to capture fine-grained musical characteristics and time-aware musical changes. To address these limitations, we propose FUTGA, a model equipped with fined-grained music understanding capabilities through learning from generative augmentation with temporal compositions. We leverage existing music caption datasets and large language models (LLMs) to synthesize fine-grained music captions with structural descriptions and time boundaries for full-length songs. Augmented by the proposed synthetic dataset, FUTGA is enabled to identify the music's temporal changes at key transition points and their musical functions, as well as generate detailed descriptions for each music segment. We further introduce a full-length music caption dataset generated by FUTGA, as the augmentation of the MusicCaps and the Song Describer datasets. We evaluate the automatically generated captions on several downstream tasks, including music generation and retrieval. The experiments demonstrate the quality of the generated captions and the better performance in various downstream tasks achieved by the proposed music captioning approach. Our code and datasets can be found in \href{https://huggingface.co/JoshuaW1997/FUTGA}{\textcolor{blue}{https://huggingface.co/JoshuaW1997/FUTGA}}.

Figures (2)

• Figure 1: Overview of FUTGA's capabilities. Given a long-form audio example, FUTGA is able to provide time-located captions by automatically detecting functional segment boundaries, as well as global captions.
• Figure 2: T-SNE representation distribution of the mixed samples of original MusicCaps songs and synthetic songs.