Analysis of unimodal Music Emotion Recognition using audio features

This repository presents different approaches to the Music Emotion Recognition task from a regression perspective, using Convolutional Neural Networks trained on MFCC audio features.
Considering the 2D representation of emotions defined by valence and arousal dimensions [Circumplex Model of Affect], two types of models are created:

• 2D-output model : predicts values for both valence and arousal
• 1D-output models : separately predict values for valence and arousal, respectively

For the complete description of the methods and experiments, please refer to unimodal MER.

Dataset

The dataset used is The MediaEval Database for Emotional Analysis of Music , consisting in 1,744 song excerpts of ~45sec duration, with two types of annotations for valence and arousal available: dynamic —measured per second— and static —measured per 45sec. In this project, the static annotations are used.

Data path structure

The data directory should have the following structure:

.
├── Data
    ├── DEAM_dataset
    │   ├── Audio
    │   │   ├── *.mp3
    │   ├── static_annotations.csv

Usage

1. Prepare data

run python main.py --mode=preprocess

Extract annotations and audio names from static_annotations.csv
Augment dataset Extract MFCC features from waveforms Make train and test sets

2. Train

run python main.py

There are three options for training:

• 2D-output model: --dimension=both will create a model to predict both valence and arousal, with filter size defined in --params_dict
• valence model: --dimension=valence will create a model to predict valence with filter size defined in --valence_params_dict
• arousal model: --dimension=arousal will create a model to predict valence with filter size defined in --arousal_params_dict

Control the training by modifying the default values for the following parameters:

--device = cuda (train on cuda)  
--log_interval = 1 (print train & validation loss each epoch)
--num_epochs = 2000

3. Test

run python main.py --mode=test --dimension=*

The model saved as Models/model_<dimension>.pt will be loaded.

• for the 2D-output model: --dimension=both
• for the valence model: --dimension=valence
• for the arousal model: --dimension=arousal

Tools

PyTorch, librosa