genre-classification

Ethan Lusterman & Robert Gruener
Department of Electrical Engineering, The Cooper Union
ECE414 Machine Learning, Fall 2013, Advisor: Sam Keene

This project collects and uses 30-second preview, genre-labeled raw audio data from echonest's API to train a prt (pattern recognition toolbox) classifier in MATLAB and then test it. Feature extraction is accomplished using Mel Frequency Cepstral Coefficients (MFCCs), and the default classifier used is a Treebagger (prtClassMatlabTreebagger). Testing of the classifier in MATLAB outputs a confusion matrix with overall percentage displayed (prtScoreConfusionMatrix). The front-end is written in Flask for uploading and testing of single audio files with visual output. Due to time constraints, the back-end uses a wrapper that allows MATLAB functions to be called from within Python (mlabwrap).

Dependencies

Python

pydub
pyechonest
numpy

MATLAB

newfolder/PRT (https://github.com/newfolder/PRT)
rastamat (http://labrosa.ee.columbia.edu/matlab/rastamat/)

Example Usage

1. Rename SAMPLE_CONFIG to CONFIG.py and add your echonest API key data.
2. python echonest.py to collect training and testing data
3. Open MATLAB and run butRunThisScriptDoe.m to load the data into MATLAB and train and test the classifier.
4. Modify the feature collection in loadData.m if desired.
5. Test classifier on single files in MATLAB using testFile.m.