In the last weeks, the UPV team has reached great advances in the automatic EDM structure segmentation. A causal low-latency algorithm has been employed for the extraction of the important events in the structure of EDM music. It employs low computational cost and does not need the full song to provide a segmentation, been appropriate for real-time performances.
We have focused our efforts on detecting events that capture the attention of the listener during playback. In this way, these moments can be associated with an important change in the video animation. The beginning and the ending of the breaks have been considered, the key events that alternate between high and low energy moments of the music. In addition, the different layers that are added or removed from the main base that build-up the song should also be detected.
To set the thresholds and fine tune our algorithm a great effort has been carried out. More than 100 songs of different styles of EDM have been selected. An expert has hand labelled every song indicating the time positions where an important event occurs. With this database, out algorithm has been trained and fine-tuned to improve its performance.
The figure above shows a typical segmentation where three different events: S = Break Start, E = Break End, L = Layer Change