Abstract
The success of computer segregation of sound sources from a single-channel mixture often relies on the estimation of multiple fundamental frequencies. Instead of solving the problem directly, this paper describes a unique audio watermarking scheme to assist sound source segregation. Individual sources are assumed to be available for watermark embedding before mixing. Each source's shorttime spectral peaks are aligned to frequency grid points labeled with binary quantization indexes. A modified sound source is synthesized with a sinusoidal model. Thus, each source is embedded with a watermark. Watermarked sources are then linearly added. To un-mix the sources, sinusoidal trajectories found in the mixture are segregated based on whether its frequency more often aligns to one or the other set of quantization grid points. Then, segregated signals are reconstructed by sinusoidal synthesis. Although the reconstructed signals sound different from the original sources, to a certain extent, the melodies can be extracted from the segregated signals by monophonic pitch estimation. © 2007 IEEE.