Evolutionary optimisation for interpretable music segmentation and music categorisation based on discretised semantic metafeatures

相似性和重复性是音乐创作的基本原则，也是形成音乐感知和分类的影响因素。该项目的任务是开发新的方法来提取音频记录的结构信息，并改善音乐的分类到流派，风格和情感。与已建立的信号分析方法相比，主要重点是提取与音乐理论相关的可解释特征。这适用于基于几个标准（乐器、和声、克里思、节奏和动态）的分段以及用于音乐分类的分段的语义特征的估计。补充目标是提高分类质量并减少对所需训练数据的需求。此外，分类模型将在鲁棒性方面进行优化（在质量降低的信号上的应用）和泛化性能（测量为来自不同类别的音乐作品的分类质量的偏差）。进化算法及其与新颖的、与问题相关的扩展的适配被选择作为用于优化特征提取、特征选择以及分段和音乐分类的多目标优化。

Similarity and repetition belong to the basic principles of music composition, but also to formative influence factors on the perception and the categorisation of music. The task of this project is to develop new methods for the extraction of structural information of audio recordings and to improve the classification of music into genres, styles, and emotions.In contrast to established signal analysis approaches, the main focus is the extraction of interpretable features which are related to music theory. This holds for the segmentation based on several criteria (instrumentation, harmony, tempo, rhythm, and dynamics) as well as for the estimation of semantic characteristics of the segments for music categorisation.The complement goals are to increase the classification quality and to reduce the demands on required training data. Besides, the classification models will be optimised with respect to the robustness (application on signals of reduced quality) and the generalisation performance (measured as the deviation of classification quality for music pieces from different categories).Evolutionary algorithms and their adaptation with novel, problem-related extensions are chosen as the core methods for the optimisation of feature extraction, feature selection, and multi-objective optimisation of segmentation and music categorisation.