Music Genre Classification from Audio Signals

• Date: 2026-01-01
• Status: Finished on 26-01-07
• Author: Colin MANYRI
• License: MIT - Copyright (c) 2026 Colin MANYRI
• Version: 10.0.0.1

Free Music Archive (FMA) – Medium Dataset

This project uses the FMA Medium (Free Music Archive) dataset, available on Kaggle.

The dataset was introduced by Michaël Defferrard, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson at the ISMIR 2017 conference. It contains over 10,000 audio excerpts of 30 seconds each, with detailed metadata allowing full traceability throughout the experimental pipeline.

• Project: Music Genre Classification
• Type: Audio signal processing & Machine Learning
• Primary Language: Python
• Project Context: Academic
• Dataset Year: 2025

Project Goal

This project aims to design and analyze a complete automatic music genre classification pipeline from raw audio signals. The focus is on mastering the entire pipeline, from loading audio files to evaluating model performance, without using precomputed features.

The project enables building and training AI models using TensorFlow for classification of the main classes of the FMA dataset. Key steps include audio preprocessing, feature extraction, TFRecord generation, baseline model definition, hyperparameter optimization via Random Search and Neuroevolution, and evaluation of the best-performing models.

Main Features

• Preprocessing and extraction of audio data (MP3 → spectrograms → TFRecord)
• Definition and training of suitable baseline models
• Hyperparameter optimization via Random Search and Neuroevolution
• Evaluation and saving of the most effective models
• Inference from a given trained model

Repository Structure

All source code is located in the src folder.
Training data is stored in the data folder, which must contain at least FMA_medium (the dataset used). The folder may also include:

• .logs: project log files
• .trash: files deleted via functions in others.py
• execution_example : show and exemple of generated folders such as optimization_results, results_quick_test, results_tests, models, containing trained models and associated information and performance analysis

Complete project documentation is located in the documents folder, including:

• model_evaluation.md: detailed reflexions on model evaluation (basis for extending the project)

• optimisation_documentation.md: explanation of model_optimization.py and optimisation_strats.py responsible for baseline model optimization

• project_structure.md: additional information on project data structures and source code structures

• research_range.md: research range defined for improving the baseline using Random Search and Neuroevolution

• strats_projet.md: planned project steps, providing insight into design thinking prior to coding

• tracks_documentation.md: header format for the tracks.csv

• Project_description.pdf: original academic project brief and grading criteria

Installation

Working Environment

• Recommended IDE: VS-Code
• AI Code Autocompletion: GitHub Copilot
• AI Code Assistance: Claude Sonnet 4.5, ChatGPT 5.1
• Python Version: 3.10
• Environment Manager: Conda

Create Virtual Environment

conda create -n FMA python=3.10
conda activate FMA

Install Dependencies

pip install -r requirements.txt

Hardware Requirements

Local AI training is resource-intensive (CPU, GPU, and RAM). Recommended minimal configuration:

• CPU: i5-12400f
• RAM: 16 GB DDR4
• GPU: RTX 4060 Ti

Data uploading

This project use external data. /!\ This repository does not contains these data.

You need to go on Kaggle to download following dataset. The minimal data requiered are :

• FMA_medium folder
• metadata folder (complete)

Please save the previous folder at the following respective path :

• {repository}/data/FMA_medium
• {repository}/data/metadata

To prevent any issue, please add these folder without with the given names in the data folder. In case of issue in the code, please check the constant folder cste.py and especialy FMA_SMALL_PATH and all constants which start with data/metadata/

External Elements / Citations

This project uses the FMA Medium (Free Music Archive) dataset, available on Kaggle : FMA - Free Music Archive - Small & Medium.

The dataset was presented by Michaël Defferrard, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson at ISMIR 2017. Copyright (c) 2016 Michaël Defferrard

Contact / Support / Author

For questions or issues regarding code execution, contact(s):

• Colin MANYRI: colin.manyri@etu.utc.fr