In recent years, the intrinsic relationship between music and motions have been widely studied. However, very few efforts have been made to develop music-driven conducting motion generation models, which takes music as input signal to generate conducting motion in harmony with music rhythm and semantics. This paper proposed a music-driven conducting motion generation approach based on DFMD. Specifically, firstly it constructed a filter using the beat information to decompose the command action into high and low frequency components. Then, a deep convolutional neural network dynamically learnt these components, and it synthesized the final command action. Experimental results on the large-scale ConductorMotion100 dataset show that the standard deviation of the generated lowfrequency and high-frequency motion components is 4.457 9 and 9.646 6, which are very close to the real motions. The proposed method breaks through the limitations of coherence and coordination in time-domain or spatial-domain motion decomposition, and effectively avoids the influence of large-value low-frequency motion on small-value high-frequency motion. The visualized results show that the generated movements are natural, beautiful, diverse, and closely synchronize with the music signal. It provides a new understanding of the connection between music and movement, and brings innovative application prospects to the field of musical performance.

[2023-12-08] Accepted Paper /doc/2023.07.0321/v1

[2024-03-05] Printed Article /doc/2023.07.0321/v2