Singing voice conversion with non-parallel data