Journal of Intelligent Computing Andre Carvalho, Hendrik Macedo 8 3 2017 http://www.dline.info/jic/fulltext/v8n3/jicv8n3_1.pdf Most of data about an item in recommendation algorithms actually consists of meta-information. We argue that meta-information heavily affects the recommendation quality. This paper analyses the impact of deeper representations and more elaborate similarity measures in the quality of the recommendation of music. Firstly, we propose a innovative approach to extract musical features from audio files and use it to deepen song representation. Next, we propose a newer metric for measuring diversification. Finally, we evaluate both diversification and accuracy for the recommendation of songs. Results show that the inclusion of these deeper features on the representation of the item improves the accuracy and greatly improves diversification.