Blind source separation is a topic of ongoing interest either as a pre-processing step for arbitrary audio analysis frameworks or for re-/up-mixing of audio streams. Many state-of-the-art algorithms are based on the non-negative tensor factorization (NTF). This thesis addresses one short-coming of the NTF: It separates only notes but not whole melodies consisting of several (different) notes of one single instrument.In this thesis, an algorithm for clustering the separated notes into melodies is developed. For this, audio features and unsupervised clustering algorithms and their strengths and weaknesses are discussed. Good pairs of audio features and clustering algorithms are shown by experiments. In order to reduce the error-rate of these clustering algorithms, strategies for combining different clustering algorithms are developed.The clustering algorithms discussed in this thesis fulfill the following requirements. They can be used unsupervised. No interaction of humans is necessary up to the signal synthesis step. Their robustness is tested on different sets of mixtures to assure the parameters to be as universally valid as possible. Finally, the proposed approach leads to comparable separation quality but can be evaluated in a fraction of the time compared to other state-of-the-art algorithms used for Blind Source Separation.