Improvement of Music Genre Classifiers Under Compressed Audio Data

We improved several music genre classifiers including support vector machines and different types of recurrent neural networks by applying light and heavy audio compression to the music GTZAN data set. First download the GTZAN dataset which contains 1000 songs split into 10 genres. Then run the compress_data.m file in the Transform Data folder that applies dynamic range compression to all the songs in the GTZAN data set. Run this twice, one for light compression (2:1 ratio) and one for heavy compression (10:1 ratio). Then we must extract features by running the files in the Feature Extraction folder. Now we run our models on the three datasets to obtain our accuracy measurements. Check out the report for more details on our findings.