Using Musical and Statistical Analysis of the Predominant Melody of the Voice to Create datasets from a Database of Popular Brazilian Hit Songs

This work deals with the creation and optimization of a large set of features extracted from a database of 882 popular brazilian hit songs and non-hit songs, from 2014 to May 2019. From this database of songs, we created four datasets of musical features. The first comprises 3215 statistical features, while the second, third and fourth are completely new, as they were formed from the predominant melody of the Voice and previously there were no similar databases available for study. The second set of data represents the graph of the time-frequency spectrogram of the singer’s voice during the first 90 seconds of each song. The third dataset results from a statistical analysis carried out on the predominant melody of the voice. The fourth is the most peculiar of all, as it results from the musical semantic analysis of the predominant melody of the voice, which allowed the construction of a table with the most frequent melodic sequences of each song. Our datasets use only Brazilian songs and focus their data on a limited and contemporary period. The idea behind these datasets is to encourage the study of Machine Learning techniques that require musical information. The extracted features can help develop new studies in Music and Computer Science in the future.