Music Generation Based on Uploaded Track List

This thesis addresses the task of music generation in the symbolic domain based on a list of uploaded tracks. The generated compositions should have a similar musical style and sound to those that were uploaded. To solve this task, a technique of selective sampling is proposed ——- a special selection of the most relevant samples of the model based on quantitative criteria. Two metrics are introduced in this work to assess the degree of stylistic similarity: a modification of the style similarity metric based on performance features and a melodic similarity metric based on the Levenshtein distance. Using the values of the proposed metrics in combination allows for the elimination of samples that are too similar to the original and the selection of the most relevant candidates from the remaining ones. The Score2Perf Music Transformer model trained on the YouTube dataset is used as the base model for sampling. The diversity of the model's stylistic distribution is demonstrated through an analysis on a sample of well-known classical pieces and a subsample of the piano-e-competition dataset. The implementation of the final music generation system in the relevant style and the conducted experiments are presented in the project repository, available at https://github.com/Vanyok-All-is-OK/music_selective_sampling.