neuronandchords/ace2017

ACE2017

Repository with scripts to reproduce some results from my 2017 MIR course student blog: https://musicinformationretrieval.wordpress.com/category/audio-chord-detection/

To run most of the scripts you'll need:

1. Linux-like or Mac OS X system with standard tools (shell, awk) and python 2.7.
2. two directories which represents chord-annotated audio dataset:
   • with 'wav' audio files (perhaps other formats (e.g. .flac) will work as well)
   • MIREX-style 'lab' files with chord annotations. Every file in one directory should be matched to a file in another directory by name (e.g.: /chains.wav corresponds to /chains.lab)
3. https://github.com/jpauwels/MusOOEvaluator compiled, with its' executable path added to the PATH environment variable.
4. Essentia (http://essentia.upf.edu/documentation/)
5. numpy python package (http://www.numpy.org)
6. vamp python package (https://pypi.python.org/pypi/vamp/1.0.0)

Additional dependencies are listed for each section.

I. State of the Art Audio Chord Estimation algorithms evaluation

(https://musicinformationretrieval.wordpress.com/2017/03/06/state-of-the-art-audio-chord-estimation-algorithms-evaluation/)

Dependencies:

1. madmom python package (https://github.com/CPJKU/madmom)
2. Essentia (http://essentia.upf.edu/documentation/)
3. Chordino vamp plugin (http://www.isophonics.net/nnls-chroma)

How to run:

1. Chordino algorithm evaluation e.g.: ./run_chordino.sh audio chordannotations
2. Madmom algorithm evaluation e.g.: ./run_madmom.sh audio chordannotations
3. Essentia ChordsDetection algorithm evaluation e.g.: ./run_essentia.sh audio chordannotations
4. Essentia ChordsDetectionBeats algorithm evaluation e.g.: ./run_essentia_beats.sh audio chordannotations

II. Robustenss tests (https://musicinformationretrieval.wordpress.com/2017/03/21/robustness-tests-for-ace-algorithms/)

NOTE: scripts for this part have hardcoded file paths which must be manually edited.

Dependencies:

1. Lame library for MP3 compression test (http://lame.sourceforge.net)
2. ffmpeg library for downsampling test (http://lame.sourceforge.net)
3. MATLAB and AudioDegradationToolbox for vynil effect test (https://code.soundsoftware.ac.uk/projects/audio-degradation-toolbox)
4. Audacity for pitch sift test (http://www.audacityteam.org)

How to run:

1. For MP3 and downsampling edit pathes and run degrader.py
2. For vynil effect run vynil.m from MATLAB
3. For pitch shift use Audacity GUI
4. For evaluation use run_on_degraded.sh, run_on_vynil.sh, run_on_shifted.sh.

III. Chroma features evaluation (http://TODO)

Dependencies:

1. for NNLS chroma: Chroma NNLS vamp plugin (http://www.isophonics.net/nnls-chroma)
2. for librosa: librosa python package (https://github.com/librosa/librosa)
3. for ChromaToolbox CLP and CRP features: MATLAB and Chroma Toolbox (http://resources.mpi-inf.mpg.de/MIR/chromatoolbox

How to run:

1. run_essentia_beats.sh for HPCP chroma features
2. run_nnls.sh for NNLS chroma
3. for CLP and CRP:
   • run extract_muller_features.m from MATLAB (manually edit hardcoded paths first).
   • ./run_clp.sh specifying obtained chroma directory and extension (either .clp or .crp).
4. for librosa: modify run_essentia_beats.sh (uncomment librosa related sections).

IV. Intermediate improvements (http://TODO)

Dependencies:

1. for NNLS chroma: Chroma NNLS vamp plugin (http://www.isophonics.net/nnls-chroma)
2. madmom python package (https://github.com/CPJKU/madmom)
3. Essentia with patched ChordsDetectionBeats algorithm (57284ab08b1bd93b24ef6b56dc7ccef5c6259d3c) should be used for second and third improvements.

Improvements:

1. NNLS chroma plugged instead of HPCP: run_nnls.sh
2. Chroma smoothed (by convolution with hanning window), sampled on every beat: run_essentia_beats_hacked.sh
3. BeatTrackerMultiFeature replaced with madmom's DBN beats: run_essentia_dbn_beats.sh
4. Template mapping pseudoprobabilities plugged into HMM: run_improved.sh

V. Final Algorithm (http://TODO)

Dependencies:

1. madmom python package (https://github.com/CPJKU/madmom)
2. Chroma NNLS vamp plugin (http://www.isophonics.net/nnls-chroma)

How to run:

1. Prepare directory with chroma and beat tracking data (in order to save time later on chord detection re-run with different parameter sets), e.g.: python prepare.py audio npz_dir
2. Run the algorithm evaluation, e.g.: ./run_improved.sh npz_dir chordannotations
3. find results in directory: IMPROVED_output_audio

VI. Plots (some of the plots used in the blog are generated with scripts from 'plot' directory)

Dependencies:

• matplotlib and seaborn python packages.