Stereo Music Source Separation for 3D Upmixing
Date:
Fri, 10/30/2009 - 2:15pm
Location:
CCRMA classroom---The Knoll, Room 217
Event Type:
DSP Seminar For this week's DSP Seminar, Hwan Shim will present his method for stereo upmixing to a binaural display (abstract follows).
Stereo Music Source Separation for 3D Upmixing
Hwan Shim , Jonathan Abel, and Koeng-Mo Sung
A method for 3D upmixing based on stereo source separation and a primary-ambient decomposition is presented. The method separately renders the primary and ambient components, and separately spatializes sources derived from the primary signal. Since all separated sources appear in the upmixed output, it is more important that the source separation method be free of audible artifacts than achieve a complete separation of the sources present. To separate source signals present in a stereo mix, the mix vector amplitude or energy is allocated to the sources present, for instance all given to the most likely source, or allocated to each source in proportion to its likelihood. However, these choices produce “musical” noise and source motion artifacts in the upmixed signal. Here, two sources are selected according to the mix vector direction, and the mix vector energy is allocated by inverting the panning matrix associated with the selected sources. Listening tests show an upmix with separated sources and few audible artifacts.
Hwan Shim , Jonathan Abel, and Koeng-Mo Sung
A method for 3D upmixing based on stereo source separation and a primary-ambient decomposition is presented. The method separately renders the primary and ambient components, and separately spatializes sources derived from the primary signal. Since all separated sources appear in the upmixed output, it is more important that the source separation method be free of audible artifacts than achieve a complete separation of the sources present. To separate source signals present in a stereo mix, the mix vector amplitude or energy is allocated to the sources present, for instance all given to the most likely source, or allocated to each source in proportion to its likelihood. However, these choices produce “musical” noise and source motion artifacts in the upmixed signal. Here, two sources are selected according to the mix vector direction, and the mix vector energy is allocated by inverting the panning matrix associated with the selected sources. Listening tests show an upmix with separated sources and few audible artifacts.