In modern music genres like Pop, Rap, Hip-Hop or Techno many songs are built in a way that a pool of small musical pieces, so called loops, are used as building blocks. These loops are usually one, two or four bars long and build the accompaniment for the lead melody or singing voice. Very often the accompanying loops can be heard solo in a song at least once. This can be used as a-priori knowledge for removing these loops from the mixture. In this paper an algorithm based on granular resynthesis and spectral subtraction is presented which makes use of this a-priori knowledge. The algorithm uses two different synthesis strategies and is capable of removing known loops from mixtures even if the loop signal contained in the mixture signal is slightly different from the solo loop signal.