Music Genre Classification using Deep learning

U ovom radu opisani su postupci svrstavanja glazbe u odgovarajuće žanrove upotrebom računala. U tu svrhu korišten je skup podataka GTZAN, koji se sastoji od 1000 primjera trajanja 30 sekundi, podijeljenih u dijelove od 5 sekundi i svrstanih u ukupno 10 žanrova. Svi postupci uključuju ručni izračun spektralnih značajki, a potom primjenu modela dubokog učenja. Od spektralnih značajki korišteni su spektrogrami STFT, Mel spektrogrami i koeficijenti MFCC. Na generiranim značajkama uspored̄eni su duboki modeli FFNN, CNN i LSTM. Najveću točnost klasifikacije od 78.83% na neviđenom skupu podataka ostvario je model LSTM koristeći značajke Mel spektrograma.This thesis describes the procedures used to utilize computers for classifying audio signals into genres. For this purpose, the GTZAN data set was used, consisting of 1000 samples with 30 seconds duration, divided into smaller chunks of 5 seconds each and labelled into a total of 10 genres. All the procedures include manual calculation of spectral features followed by the application of a deep learning model. The spectral features used are STFT spectrograms, Mel spectrograms, and MFCC coefficients. Deep FFNN, CNN and LSTM models were compared on the generated features. The highest classification accuracy of 78.83% on an unseen data set was achieved by the LSTM model using the Mel spectrogram features