Umeå universitets logga

With this thesis, we aim to automate the creation of DJ mixes. A DJ mix consists of an uninterrupted sequence of music, constructed by playing tracks one after the other, to improve the listening experience for the audience. Thus, to be able to build mixes automatically, we first need to understand the tracks we want to mix. This is done by extracting information from the audio signal. Specifically, we retrieve two pieces of information that are essential for DJs: cue points and drum transcription. In the field of music information retrieval, the two associated tasks are cue point detection and automatic drum transcription.

With cue point detection, we identify the positions in the tracks that can be used to create pleasant transitions in the mix. DJs have a good intuition on how to detect these positions. However, it is not straightforward to transform their intuition into a computer program because of the semantic gap between the two. To solve this problem we propose multiple approaches based on either expert knowledge or machine learning. Further, by interpreting the resulting models from our approaches, we also reflect on the musical content that is linked to the presence of cue points.

With automatic drum transcription, we aim to retrieve the position and the instrument of the notes played on the drumkit, to characterize the musical content of the tracks. To create the transcription, the most promising method is based on supervised deep learning. That is, models trained on labeled datasets. However, because of the difficulty of creating the annotations, the datasets available for training are usually limited in size or diversity. Thus, we propose novel methods to create better training data, either with real-world or synthetic music tracks. Further, by investigating thoroughly the performance of the models resulting from the training data, we deduce the most relevant characteristics of a dataset that help train models.

The solutions we proposed for both tasks of cue point detection and automatic drum transcription achieve high levels of accuracy. By investigating how these tasks reach this accuracy, we further our understanding of music information retrieval. And by open-sourcing our contributions, we make these findings reproducible. With the software resulting from this research, we created a proof of concept for automatic DJ mixing.