This project aims to classify electronic music. It uses libraries like librosa, sci-kit image and sci-kit learn to collect data and form models that represent songs. Currently, a Kernel Density Estimator is used to represent a genre:

House	Electro	Drum and Bass

TODO: add setup.py script

First, clone the repository and install all dependencies. To start training you must have some songs of the same genre stored in a folder like: audio/housePlaylist

Oolong's main driver is train.py

positional arguments:
    genre            classifier to train a model for
optional arguments:
    -h, --help       show this help message and exit
    -l, --load       perform analysis on the input folder
    -t, --train      train a density model on the database

Analyzying songs:

Usage: train.py genre --load folder

Running this example will create a database named with the folder name and a timestamp. Oolong will now start begin analyzing the songs in order and populate the database with their results.

This process can take a while depending on the amount of songs you are scanning in. Each song takes ~20 seconds. Each song will fill the nested class structure found in song_classes.py. The output database will be a collection of songs stored in JSON format.

Training a density model for a genre:

Usage: train.py genre --train db_path

Running this example will read in from the database JSON. Oolong will now generate a Kernel Density Estimator and display it back to the user.

Data that is larger than the average size of a song (in frames) will be thrown away. For the density calculation, a fixed x and y size is required. This is one of the core issues of this approach and is talked about in detail below.

Wait? A Kernel Density What?

A Kernel Density Estimator can be seen as a heatmap of scatterplot data. It is most comparable to a histogram in that it has bins to seperate data - however, it's goal is to not only model

This is a density model created from the features of a single song. Visually, you can see 'bins' of data forming for each kick snare pattern.

The idea is that if we combine the features of every single song's most significant segement, the output density model will be representative of the entire genre.