Using Deep Learning to Identify Multilingual Music Genres

Automatic genre classification is a widely explored topic nowadays. It is an important task as the quantity of music released is sky-rocket. A survey says that alone Spotify releases about ten thousand songs a month. At the same time, users developed a taste of listening to music of different languages. Hence, it is very crucial to maintain a data base that can automatically classify songs of any language based on their genres. This project attempts to utilize a deep learning convolutional neural network technique to identify the genre of a song from twelve different languages all over the world. This project utilized a data base of sixty songs and mainly classified them into three genres: metal, hip hop, and pop. Each song is trimmed to be exactly five seconds long, and each five second "slice" is converted into a spectrogram, a visual representation of a song’s spectrum in the time domain. These spectrograms are divided into training and testing dataset. 75% of the data set is used for training and 25% is used for testing. A tensor flow deep learning convolutional neural network is used as an image classifier to automatically segregate the songs according to their genres. This deep learning model achieved a classification accuracy of 93.3% for all of the twelve languages considered in this database.