Score-Informed Audio Parameterization of Music Signals

这个项目涉及到音乐信号自动参数化的鲁棒和实用方法的发展。特别是，目标是识别和重建与单个音符事件或整个旋律和器乐声音相对应的信号成分-这一任务与一般音频信号处理中通常称为源分离的任务密切相关。就音乐而言，一个关键的挑战是允许密切相关的音乐来源的复杂叠加，包括歌唱/器乐声音和音乐伴奏，如钢琴歌曲或歌剧。如果没有额外的知识，将一个或两个声道的音频记录分解成这样的声音几乎是不可能的。因此，在这个项目中，我们想要遵循一种知情的方法，在音频参数化过程中利用额外的分数信息。在自动计算分数-音频对齐的基础上，分数信息用于指定和指导参数化过程，并支持信号分析和重建步骤。除了对信号建模和参数优化的基础研究外，该项目还关注新应用的开发，展示其实际相关性并确保项目的可持续性。

This project deals with the development of robust and practical methods for the automated parameterization of music signals. In particular, the goal is to identify and reconstruct signal components that correspond to individual notes events or entire melodic and instrumental voices - a task being closely related to what is commonly known as source separation in general audio signal processing. In the case of music, a key challenge is to allow complex superpositions of closely related musical sources including singing/instrumental voices and accompaniments in music such as piano songs or operas. Without additional knowledge, a decomposition of a one- or two-channel audio recording into such voices is hardly possible. Therefore, in this project, we want to follow an informed approach, where additional score information is exploited in the audio parameterization process. On the basis of automatically computed score-audio alignments, the score information is used to specify and guide the parameterization process as well as to support the signal analysis and reconstruction steps. In addition to fundamental research on signal modeling and parameter optimization, this project is also concerned with the development of novel applications demonstrating the practical relevance and ensuring the sustainability of the project.