Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Mike Wu

Second Advisor

Fabio Di Troia

Third Advisor

Genya Ishigaki

Keywords

Drum Style Classification, Deep Learning, Machine Learning, Graph Neural Network, Audio Classification, Audio Spectrogram, Natural Language, Large Language Models, Music Information Retrieval, Graph Convolution Network

Abstract

Recent advances in Large Language Models (LLMs) have demonstrated outstanding capabilities in natural language understanding. However, there is still a research gap in using LLMs in the music field. This study aims to solve this challenge by establishing a systematic benchmark for LLMs on drum style classification. By using the Groove MIDI Dataset (GMD) and its expanded version (E-GMD), we investigate how different data representations can influence LLMs’ abilities to classify the drum genre. We created six data representations in music. The three representations for traditional models include image, hand-crafted features, and graph. They will be further used in Convolution Neural Networks (CNNs), Random Forest (RF), and Graph Neural Networks (GNNs), respectively. For LLMs, we transformed symbolic music into captions, events, and context-free grammar as LLMs’ input. The performance of LLMs is evaluated across zero- shot and few-shot, focusing on how examples affect the inference. To reduce the issues of data imbalance, a style merging strategy is implemented based on musical similarity. Traditional models are analyzed through accuracy, precision, recall, F1 score, weighted average, and macro average. This study provides a comprehensive evaluation of general-purpose LLMs versus task-specific models in the field of drum style classification. The results show that ResNet-50 achieves 99% accuracy, making it the best performance model, while RF can also get a similar accuracy when using data that combine MFCC and MIDI. However, when considering merely MFCC, RF produces a significant number of false negatives, showing the importance of MIDI information in drum style classification. The Graph Convolution Network (GCN) performs better than RF using pure MFCC. However, for similar styles such as FunkSoul and RockPop, GCN still cannot outperform RF with fusion data.

In contrast, LLMs perform extremely poorly compared to task-specific models without fine-tuning. The few-shot setting does not significantly improve the performance; in some circumstances, it even confuses LLMs. After applying LoRA-based fine-tuning, the Event representation improves the model accuracy to 74%, indicating that LLMs still do not match the effectiveness of traditional models in drum feature extraction.

Download

Available for download on Saturday, May 22, 2027

Share

COinS