本文为英国伦敦大学（作者：Joachim Fritsch）的硕士论文，共39页。

本硕士论文致力于解决单声道混音中音乐音源的高质量分离问题，利用乐谱中的信息来提高分离质量。这些信息被集成在音频谱图的非负矩阵分解（NMF）中，提供必要的时域和频谱信息来指导分解过程。提出了一个分解过程的总体框架，首先对合成信号学习各仪器的活化系数和基函数，然后对实际混合信号进行初始分解。将此方法应用于现有数据集，并使用BSS_EVAL和PEAS评估工具箱进行评估。然后将这些性能指标与文献中用另一种方法得到的结果进行了比较，并建立了一个新的数据集来研究各种参数对分离结果的影响。

This Master’s thesis is dedicated to the challenging problem of high quality separation of musical audio sources in monophonic mixtures, using information from musical scores to help improve the quality of separation. This information is integrated in a Nonnegative Matrix Factorization (NMF) of the audio spectrogram,providing essential temporal and spectral information to guide the decomposition process. A general framework is proposed for this decomposition process, where the activation coefficients and the basis functions of each instruments are initially learnt on synthesized signals, and then used to initialize the decomposition of the actual mixture. This method is applied to an existing dataset and assessed with the BSS_EVAL and PEASS evaluation toolboxes. The performance measures are then compared with those obtained with another method from the literature, and a new dataset is created in order to study the influence of the various parameters on the separation results.

1. 基于NMF的音乐声源分离

2. 已知评分的源分离

3. 采用本文提出方法实现的源分离

附录A MML12接收的论文

附录B TRIOS数据集

更多精彩文章请关注公众号：