Third party funded individual grant
Acronym: MU 2686/10-1
Start date : 01.10.2016
End date : 31.03.2020
Website: https://www.audiolabs-erlangen.de/fau/professor/mueller/projects/sereco
The general goal of music source separation is to decompose a music recording into its constituent signal components. One of the main problems is that the separated signals may suffer from severe audible artifacts. Considering the challenging scenario of percussive and non-harmonic sound sources, we developed in this project techniques and tools for separating and restoring drum sound events in a perceptually convincing way. We systematically approached this general source separation problem by considering a number of more specific objectives. A first goal was to develop cascaded techniques for decomposing a music mixture into (mid-level) harmonic, percussive, transient, and residual components. A second goal was to decompose drum tracks into individual drum sound events by exploiting specific properties of drum instruments. In particular, by adapting and extending Non-Negative Matrix Factor Deconvolution (NMFD) used as a central methodology of this project, we systematically studied how audio- and score-based side information can be generated, integrated, and exploited to guide the decomposition. To improve the perceptual quality of the separated drum events, a third goal was to research data-driven restoration approaches for reducing crosstalk and other undesired artifacts. Finally, we tested and evaluated our decomposition and restoration approaches by considering two different application scenarios: an audio editing application (decomposition and remixing of breakbeats) and a music analysis problem (swing ratio analysis of jazz music). Exploring novel algorithmic approaches for sound source separation within concrete application scenarios, this project contributed to fundamental research of practical relevance.
The general goal of music source separation is to decompose a music recording into its constituent signal components. One of the main problems is that the separated signals may suffer from severe audible artifacts. Considering the challenging scenario of percussive and non-harmonic sound sources, this project aims at the development of techniques and tools for separating and restoring drum sound events in a perceptually convincing way. We want to systematically approach this general source separation problem by considering a number of more specific objectives. A first goal is to develop cascaded techniques for decomposing a music mixture into (mid-level) harmonic, percussive, transient, and residual components. A second goal is to decompose drum tracks into individual drum sound events by exploiting specific properties of drum instruments. In particular, by adapting and extending Non-Negative Matrix Factor Deconvolution (NMFD) used as a central methodology of this project, we want to systematically study how audio- and score-based side information can be generated, integrated, and exploited to guide the decomposition. To improve the perceptual quality of the separated drum events, a third goal is to research data-driven restoration approaches for reducing crosstalk and other undesired artifacts. Finally, we want to test and evaluate our decomposition and restoration approaches by considering two different application scenarios: an audio editing application (decomposition and remixing of breakbeats) and a music analysis problem (swing ratio analysis of jazz music). Exploring novel algorithmic approaches for sound source separation within concrete application scenarios, this project aims at contributing to fundamental research of practical relevance.