Key-finding algorithm for symbolic and audio sources.

Project description

For v2.0.0 of justkeydding, I am planning to move entirely to Python. The main advantage is that I will be able to use the deepchroma features from madmom, which seem to deliver better results than NNLS-Chroma. Another feature in mind (WIP) is the annotation of a MusicXML|krn|MIDI file in symbolic. The code should be able to deliver an aligned-and-annotated file back (local keys as lyrics, global key annotated somewhere in the metadata) as well as the individual features. This will be released as a python package in PyPI (pip). Expected release is May 2020. Development of that new direction is happening in the pythononly branch, if you are interested in checking that code.

justkeydding

Audio and Symbolic Key Detection algorithm based on nnls-chroma and a Hidden Markov Model.

This project is the extension of a previous symbolic key detection algorithm. The original algorithm required a sequence of pitch-classes that can be easily obtained from a symbolic music file. The old project has now been integrated in this repository to deal with both, symbolic and audio files. In the audio version, the pitch-classes are obtained from chromagrams. At this moment, the algorithm supports midi, wav, and csv input files. The supported csv consist of a chromagram file as given by the Sonic Annotator program using the nnls-chroma Vamp plugin developed by Matthias Mauch and Chris Cannam.

During 2019, a few experiments showed that different key profiles worked for different types of music and different datasets, therefore, I started using the predictions from different key profiles with a meta-classifier algorithm that figured out keys better than individual key profiles did. This was also a way to train the model, as the original HMM model did not require any training from the data (and nevertheless had a good performance in MIREX ;), it was assumed that the key profiles provided the training aspect of pitch-class distributions in all keys. The current version of the model uses the ensemble method, it is slower, but hopefully, way more accurate.

Building from source

Vagrantfile

So far, this project has only been compiled and run in a virtualized Linux environment, more specifically, a Vagrant box of Ubuntu-18.04 x64. Compilation and execution in a native environment of that OS should be straighforward. I provide a simple Vagrantfile that launches the corresponding Vagrant box and installs the necessary packages to the system. We assume that Vagrant will be used to compile the project.

Cloning the repo

The project has submodules, so cloning recursively (adding the --recursive flag) is recommended:

git clone https://github.com/napulen/justkeydding.git --recursive
cd justkeydding


Launching the Vagrant box

Inside the repository, we can launch the Vagrant box

vagrant up


This takes a while, it will download the Ubuntu image (if you don't have it), spawn the vm, and compile the code. Once it finishes, the virtual machine should be running, we can then log to it

vagrant ssh


Running full algorithm

Inside the virtual machine, the repository's folder has been mounted as a shared folder in /vagrant. Go there.

cd /vagrant


Run the full algorithm (including the meta-classifier) with

./justkeydding <input_file> <output_file>


The output_file is where the key is going to be stored.

Running individual HMM

If you are interested in running the individual HMM, that is the binary located at bin/justkeydding. This program supports a few optional arguments, but in its most basic usage, it only requires an input midi|audio|chromagram_csv file as a command line argument.

bin/justkeydding <audio_filename>.wav


Acknowledgments

Thanks to Claire Arthur for providing a dataset for testing the performance of this algorithm.

Release history Release notifications | RSS feed

Uploaded source
Uploaded py2 py3