Revisiting Autotagging Toward Faultless Instrumental Playlists Generation

This study deals with the classification of Instrumentals and Songs in a bigger musical database than what was used in all previous studies. Songs are musical pieces containing singing voice, contrary to Instrumentals. This research tackles the imbalance between the number of Instrumentals and the numerous Songs present in industrial musical databases. Our work considers the low precision of automatically generated playlists from content-based audio retrieval. Indeed, previous works failed to address an efficient Instrumental detection algorithm. We set up three experiments to assess the flaws of previous works on an original and bigger musical database. This paper posits a new approach that uses the presence probability of a frame's predicted singing voice to deduce the track tag, i.e., whether the track is an Instrumental or a Song. The main novelties are twofold. Firstly, we propose two sets of features at the track scale based on frame predictions. Secondly, we propose a paradigm that focuses on minority classes in musical databases and thus enhances general user satisfaction. The suggested approach has a better Instrumental detection. Thus, it can be used to generate thematic playlists with up to three times less false positives than existing playlists. Furthermore, we provide the source code to guarantee reproducible research. We also propose clues for further research toward faultless playlists for other tags.