Recently, our institute constructed a working prototype of a drum "robot", that is able to accompany a musician by "listening" to what he/she is playing, detecting the rhythmic structure, and playing accordingly. However, it is limited in what it can do:

• It can only detect a small number of rhythms.
• It does not handle rhythm variations very well.
• For each of these rhythms, it can only play a predefined drum pattern.
• Learning new rhythms is a manual, tedious task.

This context opens the possibility to work on multitude of topics. These topics range from practical implementations of algorithms with real-time constraints, to applied (offline) machine learning.

Note that musical knowledge is not strictly necessary to work on any of these topics.

Open Topics:

• Finding Suitable Drum Patterns for Given Rhythmic Inputs
  without the need to manually specify drum patterns for each learned rhythm. Here, we assume that we have a database of drum rhythms to choose from.
• Generating Drum Patterns (Fitting Certain Styles or Rhythms).
  The goal is to create a generative model that is able to create new, good sounding drum rhythms on demand.
• Interactive Ad-hoc Learning of New Rhythms.
  Simplify the task of "teaching" Drum-O-Tron new rhythms by leveraging the power of transfer learning methods. The goal is to enable the user to conveniently teach the computer a new rhythm giving only few examples.
• Real-time Downbeat Tracking Robust to Rhythmic Variations.
  Make Drum-O-Tron less sensitive to rhythmic variations by either inventing a new real-time downbeat tracker or combining existing approaches available at our institute.
• Automatic Drumstick Delay Calibration.
  Before using Drum-O-Tron, it is necessary to configure the delay compensation for each stick individually so Drum-O-Tron can play in time. Implement and test a method that automatically determines the delay and configures Drum-O-Tron accordingly.

Contact: Florian Krebs, Filip Korzeniowski