Genre Classification with Deep Learning Techniques

CARNAGHI, MARCO; CEBEDIO, CELESTE

La Plata

Congreso; Argentine Conference on Embedded Systems; 2022

Universidad Nacional de La Plata

Music recommendation systems are aimed at improve the listening and search experience of music consumers. The increased access to digital content has turned the algorithmic recommendation systems into a necessity in order to save the time of the users.A common approach is to employ metadata obtained from the audio track to generate the different recommendations. Music genre is one of the most important descriptors in the decision process.Therefore, in this paper, the performance of three neural networks architectures for automatic genre classification is compared. The dataset, preprocessing of audio data, training process and concepts employed for models combination is also presented.The results indicate that to combine the local pattern representation ability of CNN models with the time relationship overview of RNN models leads to an improvement in the performance of the system.The proposed models and their analysis focus on improving classification accuracy while maintaining a reduced number of layers to allow an easy implementation, whether in a embedded systems or as flexible module within another application.