Informed Source Separation for Multiple Instruments of Similar Timbre

This Master’s thesis focuses on the challenging task of separating the musical audio sources with instruments of similar timbre. We address the case in which external pitch information to assist the separation process is available. This information is provided to the source / filter model, which is embedded in a Non-Negative Matrix Factorization (NMF) framework that processes the audio input spectrogram. Different state of the art literature methods are inspected and extended. As an extension to these, two new separation methods are proposed, the Multi-Excitation and Single Filter Instantaneous Mixture Model and the Multi-Excitation and Multi-Filter Instantaneous Mixture Model. The use of dedicated source and filter decomposition for each instrument is proposed. In addition, we introduce the use of timbre models in the separation process. Timbre models are previously trained on isolated instrument recordings. The methods are compared with the BSS Eval and PEASS evaluation toolkits over an existing dataset. Promising results obtained in the conducted experiments, which shows that this is a path to be further investigated.