Generating Music for Video Games with Real-Time Adaptation to Gameplay Pace

The paper aimed to develop an automatic music generation method for video games that could create various types of music to enhance player immersion and experience, while also being a more cost-effective alternative to human-composed music. The issue of automatic music generation is an interdisciplinary research area that combines topics such as artificial intelligence, music theory, art history, sound engineering, signal processing, and psychology. As a result, the literature to review is vast. Musical compositions can be analyzed from various perspectives, and their applications are extensive, including films, games, and advertisements. Specific methods of music generation may perform better only in a narrow field and range, although we rarely encounter fully computer-generated music nowadays. The proposed algorithm, which utilizes RNN and 4 parameters to control the generation process, was implemented using PyTorch and real-time communication with the game was established using the WebSocket protocol. The algorithm was tested with 14 players who played four levels, each with different background music that was either composed or live-generated. The results showed that the generated music was enjoyed more by the players than the composed music. After implementing improvements from the first round of play tests, all of the generated music received better evaluations than the composed looped music in the second run.