Intimate Musical Collaboration with a Probabilistic Model
Karl Johannsson, Thor Magnusson, Victor Shepardson · Zenodo (CERN European Organization for Nuclear Research) · 2023
https://aimc2023.pubpub.org/pub/v44bkgoy Recent advancements in deep learning have created many opportunities in the field of music. This research explores training a deep learning model to play a percussion instrument collaboratively with a human player. The aim is to create a system that convincingly responds to the human player’s performance in real-time. Music generation is more commonly handled offline than in a performance environment, where the model’s response must feel instantaneous, organic, and sensible in response to the user’s input. A probabilistic model was trained on two large datasets, and transfer learning used to train the model on high quality data abstracted from a collaborative percussion instrument, the txalaparta. The paper outlines the design and implementation of the system, including data collection, model training, and performance visualization. The research contributes to the field of interactive music systems and demonstrates the potential of deep learning in creating intelligent musical systems that can collaborate with human performers.