Audio Classifier

Project Overview

The Audio Classifier project is a machine learning pipeline designed to classify audio tracks into various genres. It leverages the Free Music Archive (FMA) dataset and processes audio files into log-mel spectrograms for training and evaluation. The project is still under development, with additional scripts and features planned for future implementation.

Features

• Audio Preprocessing: Converts raw audio files into log-mel spectrograms using librosa and saves them as PyTorch tensors.
• Custom Dataset: A PyTorch Dataset class for loading spectrograms and their corresponding labels.
• Metadata Handling: Reads and processes metadata from the FMA dataset.

Installation

Prerequisites

• Python 3.13 or higher
• Linux operating system
• uv package manager (for dependency management)

Dependencies

Install the required dependencies using the following command:

uv pip install -r pyproject.toml

Or, to add a new dependency:

uv pip install <package-name>

All dependencies are managed via pyproject.toml and uv.lock. You do not need a requirements.txt file.

Usage

Preprocessing Audio Files

To preprocess audio files and generate log-mel spectrograms, run the preprocess.py script:

uv run python -m scripts.preprocess

This script processes audio files in the data/raw/fma_small directory and saves the spectrograms in the data/processed directory.

Loading the Dataset

The AudioDataset class in src/data_loader.py can be used to load the processed spectrograms and their corresponding labels:

from src.data_loader import AudioDataset

dataset = AudioDataset(
    processed_data_dir='data/processed',
    metadata_file='data/fma_metadata/tracks.csv'
)
print(f"Number of samples: {len(dataset)}")
spec, label = dataset[0]
print(f"Sample spec shape: {spec.shape}, Sample label: {label}")

Project Structure

audio_classifier
├── main.py
├── pyproject.toml
├── README.md
├── uv.lock
├── data/
│   ├── fma_metadata.zip
│   ├── fma_metadata/
│   │   ├── tracks.csv
│   │   └── ...
│   ├── processed/
│   │   ├── 000002.pt
│   │   └── ...
│   └── raw/
├── scripts/
│   └── preprocess.py
└── src/
    ├── data_loader.py
    ├── utils.py
    └── ...

Future Work

• Implement additional scripts for model training and evaluation.
• Add support for more audio formats and datasets.
• Develop a FastAPI-based web interface for real-time audio classification.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Acknowledgments

• The Free Music Archive (FMA) dataset for providing the audio tracks and metadata.
• The developers of librosa, PyTorch, and other open-source libraries used in this project.