Starting off as a mini-project for a class, it turned out to be a project to help me get learn and explore Weka, a data mining framework and using MEX functions in Matlab.
I used the GZTAN genre dataset as my source audio, extracting multiple low level features like Spectral Centroid, Spectral Decrease and Mel Frequency Cepstral Coeffecients and deriving various sub-features from the initial set.
After exporting the features from Matlab into Weka, I standardized the inputs, and tested different classifiers. I had about 76% correctly classified instances with the libSVM wrapper, and about 58.6% the C4.5 algorithm (J48 on Weka). A Random Forest with 10 trees gave me a rate of 65%, while forest with 20 trees gave me about 72%.
I’m using Weka for my Master’s project, identifying bird species through audio. Will post updates soon!Tags: 2014, Audio Content Analysis, Georgia Tech, Machine Learning, Matlab, Weka