Code for the music autotagging algorithm developed during my PhD at the Music Technology Group of the Universitat Pompeu Fabra.

This algorithm was submitted to the MIREX competition in 2011 in the Audio Tag Classification Task and obtained the 3rd best result.

The code is licensed under a GPLv3 license, so feel free to use it, modify it and redistribute it as long as you respect the license :-)

For researchers please cite the following papers if you use this code:

@article{sordo2012semantic,
  title={Semantic annotation of music collections: a computational approach},
  author={Sordo, Mohamed},
  year={2012},
  publisher={Universitat Pompeu Fabra}
}

or

@inproceedings {sordo20071,
 author = {Sordo, M. and Laurier, C. and Celma, O.},
 title = {Annotating Music Collections: How content-based similarity helps to propagate labels},
 booktitle = {Proceedings of 8th International Conference on Music Information Retrieval},
 year = {2007},
 address = {Vienna, Austria},
}

If you use the sample dataset included in this repository (crawled from the majorminer website), please cite the following paper:

@article{mandel08b,
  title = {A Web-Based Game for Collecting Music Metadata},
  author = {Michael I. Mandel and Daniel P. W. Ellis},
  journal = {Journal of New Music Research},
  year = {2008},
  volume = {37},
  number = {2},
  pages = {151--165},
  url = {http://m.mr-pc.org/work/jnmr08.pdf},
}

The two main dependencies of this code are Essentia and Gaia, two libraries developed at the Music Technology Group of the Universitat Pompeu Fabra. Please check their respective websites for more information. In a nutshell, this code uses Essentia to extract audio features for a collection of songs, and Gaia to represent those audio features in a high dimensional space, and perform operations such as dimension reduction, feature selection and similarity measurement.

In my case, I follow these steps in a Linux Ubuntu distribution:

• sudo apt-get install build-essential libqt4-dev libyaml-dev swig python-dev pkg-config
• git clone https://github.com/MTG/gaia.git
• cd gaia/
• ./waf configure --download --with-python-bindings
• ./waf
• sudo ./waf install

• Download essentia v. 2.0.1. This version includes pre-trained high-level classification models for genres, moods, rhythm and instrumentation. More info here

• Run:

  • sudo apt-get install build-essential libyaml-dev libfftw3-dev libavcodec-dev libavformat-dev python-dev libsamplerate0-dev libtag1-dev
  • sudo apt-get install python-numpy-dev python-numpy
  • (optional, only if './waf' fails loading liblapack) sudo apt-get install --reinstall libatlas-base-dev
  • tar xzf v2.0.1.tar.gz

• Edit the streaming extractor example to output yaml files instead of json (needed by Gaia)

  • open file 'essentia-2.0.1/src/examples/streaming_extractor_archivemusic.cpp' and modify line 127:
    • outputToFile(stats, outputFilename, true); --> outputToFile(stats, outputFilename, false);

• Now go to the essentia folder ('essentia-2.0.1/') and run:

  • ./waf configure --mode=release --with-python --with-cpptests --with-examples --with-vamp
  • ./waf
  • (optional) sudo ./waf install

• Finally, copy or create a symbolic link to the pre-trained svm models, for example:

  • ln -s essentia-2.0.1/src/examples/svm_models/

• sudo pip install unipath

The code is basically divided in 5 main blocks:

• Feature extraction
• Dataset creation
• Feature selection
• Training and Classification (Autotagging)
• Evaluation

There is an example.py script that shows how to perform autotagging over a sample dataset (crawled from the majorminer website), using k-fold cross validation. This script also calculates all the evaluation measures used in the MIREX competition. To run it, you just need to change a couple of paths in the feature_extraction function and the path to the features in the metadata file.

Each block may also be run separately. Each block has a running script (feature_extraction.py, dataset_creation.py, feature_selection.py, autotagging.py) with different options to customize the algorithm.

####Note

• The metadata files should follow this format:
  • trackid[TAB]trackpath[TAB]tag([TAB]weight)
• The feature files should follow this format:
  • trackid[TAB]track_features_path

This is already explained in the code.

If you have any doubt, please feel free to contact me.