This repository contains the code accompanying the paper "Towards Unified Music Emotion Recognition across Dimensional and Categorical Models" by Dr. Jaeyong Kang and Prof. Dorien Herremans.

🔥 Live demo available on HuggingFace

Introduction

We present a unified multitask learning framework for Music Emotion Recognition (MER) that integrates categorical and dimensional emotion labels, enabling training across multiple datasets. Our approach combines musical features (key and chords) with MERT embeddings and employs knowledge distillation to enhance generalization. Evaluated on MTG-Jamendo, DEAM, PMEmo, and EmoMusic, our model outperforms state-of-the-art methods, including the best-performing model from the MediaEval 2021 competition.

Change Log

• 2025-04-16:

  • Improved handling of short audio files (under 10 seconds) in the Hugging Face demo.
  • Fixed edge case bugs in chord processing (e.g., unrecognized "X" chord root).

• 2025-02-10:

  • Released Music2Emo v1.0, featuring both categorical and VA emotion prediction from music.

Quickstart Guide

Predict emotion from audio:

from music2emo import Music2emo

input_audio = "inference/input/test.mp3"

music2emo = Music2emo()
output_dic = music2emo.predict(input_audio)

valence = output_dic["valence"]
arousal = output_dic["arousal"]
predicted_moods =output_dic["predicted_moods"]

print("\n🎵 **Music Emotion Recognition Results** 🎵")
print("-" * 50)
print(f"🎭 **Predicted Mood Tags:** {', '.join(predicted_moods) if predicted_moods else 'None'}")
print(f"💖 **Valence:** {valence:.2f} (Scale: 1-9)")
print(f"⚡ **Arousal:** {arousal:.2f} (Scale: 1-9)")
print("-" * 50)

Installation

This repo is developed using python version 3.10

git clone https://github.com/AMAAI-Lab/Music2Emotion
cd Music2Emotion
pip install -r requirements.txt

• Our code is built on pytorch version 2.3.1 (torch==2.3.1 in the requirements.txt). But you might need to choose the correct version of torch based on your CUDA version

Dataset

Download the following datasets:

• MTG-Jamendo (Link)
• PMEmo (Link)
• DEAM (Link)
• EmoMusic (Link)

After downloading, place all .mp3 files into the following directory structure:

dataset/
├── jamendo/
│   └── mp3/**/*.mp3    # MTG-Jamendo audio files (nested structure)
├── pmemo/
│   └── mp3/*.mp3       # PMEmo audio files
├── deam/
│   └── mp3/*.mp3       # DEAM audio files
└── emomusic/
    └── mp3/*.mp3       # EmoMusic audio files

Directory Structure

• config/: Configuration files
• dataset/: Dataset directories
• dataset_loader/: Dataset loading utilities
• utils/: Other utilities
• model/
  • linear.py: Fully connected (FC) layer with MERT features
  • linear_attn_ck.py: FC layer with MERT and musical features (chord/key)
  • linear_mt_attn_ck.py: Multitask FC layer with MERT and musical features (chord/key)
• preprocess/
  • feature_extractor.py: MERT feature extraction
• saved_models/: Saved model weight files
• data_loader.py: Data loading script
• train.py: Training script
• test.py: Testing script
• trainer.py: Training pipeline script
• inference.py: Inference script
• music2emo.py: Video2Music module that outputs emotion from input audio
• demo.ipynb: Jupyter notebook for Quickstart Guide

Training

  python train.py

Test

  python test.py

Evaluation

Comparison of performance metrics when training on multiple datasets.

Comparison of our proposed model with existing models on MTG-Jamendo dataset.

TODO

• Incorporate additional features, such as lyrics.

Citation

If you find this resource useful, please cite the original work:

@misc{kang2025unifiedmusicemotionrecognition,
      title={Towards Unified Music Emotion Recognition across Dimensional and Categorical Models}, 
      author={Jaeyong Kang and Dorien Herremans},
      year={2025},
      eprint={2502.03979},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2502.03979}, 
}

Kang, J. & Herremans, D. (2025). Towards Unified Music Emotion Recognition across Dimensional and Categorical Models, arXiv.