Synergies between Musical Source Separation and Instrument Recognition

Due to the increasing amount of digital music available, there is a clear need of a proper organization and effective retrieval. Automatic instrument recognition techniques are useful for satisfying such needs, by labeling music pieces with their instrumentation, but also as support to the extraction of other semantic information such as the genre. Source separation has also recently been applied to facilitate the analysis of musical data, as well as to other applications such as karaoke or post production. In contrast with the huge need for both algorithms, the results obtained so far show that there is still much room for improvement. The main purpose of this thesis is to find synergies between instrument recognition and source separation algorithms in two different tasks: 1) the separation of a target instrument from the accompaniment, and 2) the automatic labeling of songs with the predominant music instruments. Several combination strategies are presented, aimed at overcoming some of the limitations of current state-of-the-art algorithms. In the first task, instrument recognition is used to detect the presence of the target instrument in order to apply or bypass the separation algorithms. In the second task, source separation is used to divide the polyphonic audio signal into several streams, given as input to the instrument recognition models. Promising results were obtained in the conducted experiments, showing that this is a path to be further investigated.