Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Training Music Taggers on Synthetic Data

Nadine Kroher, Steven Manangu, Aggelos Pikrakis

TL;DR

This study investigates training music taggers with synthetic data using GTZAN-synth, a tenfold larger synthetic analogue produced with MusicGen guided by LLM-driven prompts. It compares naive data augmentation against domain adaptation, transfer learning, and fine-tuning to bridge real–synthetic distribution gaps. Key findings show that simply mixing synthetic data with real GTZAN provides no real-data gains, while transfer learning and fine-tuning from synthetic pretraining offer modest improvements; domain adaptation reduces representation mismatch but yields limited accuracy gains. By releasing GTZAN-synth and outlining practical strategies, the work lays groundwork for future research on prompt engineering, scaling synthetic data, and applying these methods to broader music tagging tasks.

Abstract

Most contemporary music tagging systems rely on large volumes of annotated data. As an alternative, we investigate the extent to which synthetically generated music excerpts can improve tagging systems when only small annotated collections are available. To this end, we release GTZAN-synth, a synthetic dataset that follows the taxonomy of the well-known GTZAN dataset while being ten times larger in data volume. We first observe that simply adding this synthetic dataset to the training split of GTZAN does not result into performance improvements. We then proceed to investigating domain adaptation, transfer learning and fine-tuning strategies for the task at hand and draw the conclusion that the last two options yield an increase in accuracy. Overall, the proposed approach can be considered as a first guide in a promising field for future research.

Towards Training Music Taggers on Synthetic Data

TL;DR

This study investigates training music taggers with synthetic data using GTZAN-synth, a tenfold larger synthetic analogue produced with MusicGen guided by LLM-driven prompts. It compares naive data augmentation against domain adaptation, transfer learning, and fine-tuning to bridge real–synthetic distribution gaps. Key findings show that simply mixing synthetic data with real GTZAN provides no real-data gains, while transfer learning and fine-tuning from synthetic pretraining offer modest improvements; domain adaptation reduces representation mismatch but yields limited accuracy gains. By releasing GTZAN-synth and outlining practical strategies, the work lays groundwork for future research on prompt engineering, scaling synthetic data, and applying these methods to broader music tagging tasks.

Abstract

Most contemporary music tagging systems rely on large volumes of annotated data. As an alternative, we investigate the extent to which synthetically generated music excerpts can improve tagging systems when only small annotated collections are available. To this end, we release GTZAN-synth, a synthetic dataset that follows the taxonomy of the well-known GTZAN dataset while being ten times larger in data volume. We first observe that simply adding this synthetic dataset to the training split of GTZAN does not result into performance improvements. We then proceed to investigating domain adaptation, transfer learning and fine-tuning strategies for the task at hand and draw the conclusion that the last two options yield an increase in accuracy. Overall, the proposed approach can be considered as a first guide in a promising field for future research.
Paper Structure (15 sections, 2 equations, 3 figures, 1 table)

This paper contains 15 sections, 2 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related work
  3. Music tagging
  4. Generative music systems
  5. Domain adaptation
  6. Data
  7. Real music dataset
  8. Synthetic music dataset
  9. LLM-prompt engineering
  10. MusicGen-prompt engineering
  11. Music generation
  12. Model Architecture
  13. Supervised domain adaptation
  14. Experimental evaluation
  15. Conclusions

Figures (3)

  • Figure 1: Schematic representation of the proposed prompt engineering pipeline. For a given music genre, we use an LLM to generate genre-specific track descriptions. After some additional processing, these serve as input to MusicGen, a model for text-conditioned music generation.
  • Figure 2: Schematic representation of the model architecture which takes a mel-spectrogram as input and predicts the music genre in a multi-class classification task. 2D-Conv: two-dimensional convolutional layer. LN: layer normalization. CC: concatenation. MaxPool: maximum pooling. AvgPool: average pooling.
  • Figure 3: t-sne visualisations of intermediate representations of a subset of real and synthetic data with (bottom) and without (top) DA for three genres.