little side project: code to learn ballroom dance style from audio goals: -web page scraping -data gathering from the web -base line check -getting aquainted with keras -evaluate different ml/deep models and audio preprocessing techniques

training data downloaded from youtube labels from www.tanzmusik-online.de

current caveeats: songs with multiple assigned labels are added to data of both labels; -> should have an implementation that displays likelihood of different labels and compares to all possible labels

according to https://mediatum.ub.tum.de/doc/1138535/582915.pdf# classification rate should be very good -> more suitable audio features? use the ones from the paper;

no hyperparameter search

training is memory limited -> use batch wise data generator from keras

models in use: RandomForestClassifier (as baseline) NN RNN LSTM CNN

models to be implemented multi-label SVM other models with multi-label prediction -> requires other means of assessment

audio preprocessing in use: raw rectified "emg" features mfcc & filter banks

models to be implemented audio features like beats per minute (https://mediatum.ub.tum.de/doc/1138535/582915.pdf#)

depends on: bs4.BeautifulSoup keras numpy os.path pandas scipy.io scipy.fftpack sklearn tensorflow time urllib2